🧠 Verimem

Verified memory for AI agents · verimem.com (engine formerly known as Engram / HippoAgent — "engram" remains the architecture term: the inspectable memory-trace artifacts this system produces)

Persistent memory layer for Claude Code (and any LLM agent). Most AI agents have amnesia — and the ones with memory will store anything, true or not. Verimem screens every write with a lexical anti-confabulation filter by default, and — when you pass a source and ground=True — admits the fact only if that source actually entails it (the AUROC-0.971 moat no competitor ships). Reads carry provenance. A hippocampus with a notary at the door.

📊 Live state — run engram status for the current counts (they grow as you use it). This repo's corpus is ~550 episodes · ~325 skills · ~4.6k facts · 231 MCP tools at the time of writing — see STATE.md for the honest snapshot (what works, what doesn't, and the roadmap).

Why the engram architecture? In neuroscience, an engram is the physical substrate of a memory trace — the actual change in neural tissue that encodes a learned experience (the term was coined by Richard Semon in 1904; Tonegawa's lab at MIT later gave it physical evidence through optogenetic engram cell studies). That's exactly what this system produces: not weights, but inspectable, fitness-tracked, mergeable memory artifacts that persist across sessions.

Verimem is not a replacement for Claude Code, Cursor, or your favourite agent. It's the memory module they're missing. Plug it in via MCP and:

• Episodes you and the agent worked on persist across sessions.
• Procedures the agent figured out get compiled into deterministic macros — the next call doesn't even hit the LLM.
• Failure → success patterns are rescued and replayed during sleep cycles.
• Skill consolidation runs subscription-first (cycles #34-#40 Hippo Dreams): the host's LLM does the heavy lifting via Pro/Max plan — zero extra API spend, atomic shadow → review → adopt with rollback on failure.

8+ neuroscientific mechanisms (DG pattern separation, TCM, synaptic tagging, lateral inhibition, engram crossover, schema priming, …) all opt-in via flags.

🏆 Retrieval moat — measured, not claimed

Verimem's recall fuses dense-cosine + entity-PPR + BM25-lexical via RRF, then a cross-encoder rerank — a combination no competitor ships (HippoRAG-2: no CE, no lexical; Zep: no CE; Mem0: cosine only). On LongMemEval-s (public benchmark, session-level retrieval, judge-free, the same e5-base embedder for every arm, zero external APIs):

arm (FULL 500, like-for-like)	recall@5
Verimem, fusion OFF (bare cosine)	0.8525
Verimem, fusion ON (3-signal)	0.8745

+2.2 pp from the fusion on the full 500 (A/B toggling only ENGRAM_PPR_FUSION; lme_s_fusionON/OFF_n500_clean.json). An earlier n=300 sample read 0.909/+7.5pp, but it covered only 4 of the 6 question types and under-sampled temporal-reasoning (the weak type at 0.793) — the full-500 number above is the honest one; the n=300 headline was optimistic. mem0 2.0.4 (e5-parity) sat at 0.790 on an n=100 slice (not full-500, so not strictly comparable). The stage-2 cross-encoder rerank does not move LongMemEval recall@5 (it reorders within the top-k; the gold session is already in top-5) — its verified lift is R@1 on hard fact paraphrases, a different task. Fusion is Default-ON; opt-out with ENGRAM_PPR_FUSION=0. On top of retrieval, Verimem ships two things the others don't: bi-temporal valid-time (valid_until hard-expire — the differentiator vs Mem0/Zep) and a write-time prompt-injection screen that is multilingual (incl. Italian) — mem0 and engram-memory ship none.

On the write path, a tunable trust-maintenance dial no competitor ships: anti-sycophancy (a bare assertion can't overwrite a verified fact), evidence-tiered supersession, and a precision floor — recall-first ↔ never-delete-truth, all env-driven and default-off. Numbers, knobs and honest caveats: docs/TRUST_MAINTENANCE.md.

On the read path, the same philosophy as a measured 4-point dial (HaluMem QA, n=120, like-for-like — competitors ship one unmeasured operating point):

arm	correct	hallucination
baseline	0.408	0.233
ENGRAM_RECALL_CENTERING=1	0.433	0.225
ENGRAM_GROUNDING_GATE=1	0.358	0.125
both	0.350	0.111

Every answer is externally verified against the retrieved evidence; unsupported answers become abstentions — hallucinations −52% at the trust-first end, your call where to sit. And it holds at scale: brute-force recall is O(N) (81 ms @1M) while the wired ANN stays ~flat (1.3 ms) — 62× @1M, sublinear (reproducible): SCALE.md.

Honest scorecard vs MemOS (self-reported) on HaluMem: extraction F1 0.71–0.74 vs 79.7 (gap −6/−9pp, was −14.7); updating 0.25–0.29 judged vs 62.1 — retrieval-capped (4 embedders across 3 families within ~3pp; the paraphrase-matching ceiling is a model-class limit we state instead of hiding); QA end-to-end 0.553 vs 0.672 (behind; extraction recall is the bottleneck — the same recipe on a gold store reads 0.739–0.750, split detailed in the comparison section below) with the hallucination trade above. Our axis is measured trust + reproducible scale + the dial — not raw-recall parity.

Full numbers, fairness notes and honest limits: BENCHMARKS.md.

⚡ Install in 2 minutes (Claude Code)

pip install verimem
# From source (latest main):  pip install "git+https://github.com/aureliocpr-ctrl/verimem.git"
# For development:  git clone https://github.com/aureliocpr-ctrl/verimem && cd verimem && pip install -e .

Drop .mcp.json in your project root (or ~/.claude/.mcp.json global):

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "args": ["mcp"],
      "env": {
        "ENGRAM_HOSTED": "1",
        "ENGRAM_DATA_DIR": "${HOME}/.engram"
      }
    }
  }
}

Restart Claude Code. The 231 hippo_* MCP tools become callable; the hippoagent-memory skill auto-activates at every session. To disable temporarily: export ENGRAM_DISABLED=1 (legacy HIPPO_DISABLED=1 still works).

First run downloads a ~440 MB embedding model (multilingual-e5-base). Pre-download it once so your first recall is instant instead of waiting on a silent background fetch:

engram warmup        # downloads + loads the model, reports when ready

It's also the natural pre-bake step for CI / Docker images.

Use it from Python (no MCP, 3 lines)

from engram import Memory

mem = Memory()                                          # local SQLite, offline
mem.add("The deployment uses PostgreSQL 16.")           # write goes THROUGH the gate
for hit in mem.search("which database?"):               # read returns provenance
    print(hit["text"], hit["status"], hit["grounding_score"])

Same add() / search() ergonomics as mem0/Zep — with the difference that is the whole point: add() routes every write through the anti-confabulation gate (lexical screen + optional contradiction + optional source⊢fact entailment), so an unsupported claim is downgraded or refused instead of stored; and search() returns each fact's provenance (status, write-time grounding_score) so your code can trust-condition. Pass source= and enable ENGRAM_GROUNDING_WRITE for the strongest gate (source-entailment, AUROC 0.971).

Full surface (mem0/Zep parity): add · search/recall · get · get_all · update · delete, plus the one a cosine-only store can't offer — history(id), the supersession chain (update() never destroys the old fact, it supersedes it, so the provenance trail is auditable):

mem.update(fid, "The server now listens on port 9090.")  # supersedes, keeps the old
for v in mem.history(fid):                                # oldest -> newest
    print(v["text"], "->", v["superseded_by"])

Make it fully automatic (recommended)

Add the Memory Protocol to your ~/.claude/CLAUDE.md so Claude writes to memory at end of every significant task and reads at start — no slash commands, no friction. Two-line setup, details here.

For explicit control instead, install the optional slash commands from slash_commands/ (/recall, /remember, /memory, /forget, /skill-top).

Real-world use cases

• General developer workflow → docs/MEMORY_PROTOCOL.md
• Autonomous pentesting agent (NEXUS) → docs/usecases/pentesting_nexus.md

📊 How Verimem compares (honest, no spin)

HaluMem QA — the composed trust recipe, adversarially reviewed (2026-07-06)

With the full recipe ON (bi-temporal asserted_at + reconcile floor 0.35 + verification-aware answerer + dated history context + k=12), per-session QA on HaluMem-Medium (u0 evidence):

Two DIFFERENT measurements, never on the same row (adversarial review C8 — an earlier draft compared our read-path number against MemOS's end-to-end number without saying so):

	user 1 (n=188)	user 0 (n=164)	user 2 (n=169)	MemOS end-to-end (self-reported)
Read-path QA — store built from gold memory points; measures the memory (recall+history+verify+abstention), not our extractor	0.739	0.750	0.787	—
End-to-end QA — our extraction → gated store → answer, same questions (run 1, run 2)	0.553* (×2 runs)	not run	not run	0.672
Memory Boundary, never fabricate (read-path / e2e)	0.976 / 1.000	0.897 / —	1.000 / —	—
Memory Conflict (read-path / e2e)	0.800 / 0.800	0.872 / —	0.889 / —	—
Basic Fact Recall (read-path / e2e)	0.800 / 0.267	0.725 / —	0.750 / —	—

Read-path across 3 users: 0.739 / 0.750 / 0.787 (mean 0.759) — every run above MemOS's end-to-end 0.672 (u2 evidence).

End-to-end we are BEHIND MemOS (−12pp) — stated, not hidden. The split shows where: the trust properties HOLD through the full pipeline (Boundary 1.000 on both runs, Conflict 0.70–0.80) while extraction recall on basic facts is the bottleneck (0.27–0.38 e2e vs 0.800 read-path). *Two independent runs both scored 0.5532 (stable overall). Between them we found and fixed an identity-leak artefact by reading run-1's failures (the extractor baptised anonymous speakers with an out-of-text name — reproduced live, fixed): run 2 has 3/188 leak-tainted answers vs 13/188, and Basic Fact +11pp where the fix bites; per-category shifts elsewhere are within run-to-run variance.

The read-path claim survived an adversarial counterexample review (claim_holds, conf 0.86: per-category counts consistent by construction; abstention flagging doesn't inflate — fabrications are penalized; answerable categories graded by the strict judge). Declared caveats, every time: judge is Claude (MemOS self-reports with GPT-4 — comparable method, not identical); n=3 users; the rich context trades a little abstention purity (0.976/0.897 vs 1.0 with the strict answerer). The single biggest lever is the answer-with-history dated context: on transition questions it lifts +16pp (7 unlocked, 0 lost), and the Memory-Conflict arc is 0.15 → 0.825 (5.5×) from plain store to full recipe — see TRUST_MAINTENANCE.md for the measured dial and QUICKSTART_SDK.md for the 5-verb SDK that ships it (add/search/history/as_of/explain — archaeology, time-travel and the evidence dossier included). Write-path robustness, same block: reconcile auto-supersede made surgical (146/807 retired, 0 cross-attribute, 3× faster) and a 7/7 stress battery (evidence).

TrustMem-Bench — the trust benchmark we impose (offline, run it yourself)

Every memory benchmark measures accuracy. None measures whether a memory can be trusted — so we built one. Seeded synthetic personas (EN + IT), six axes whose verdict is deterministic (no LLM, no network), one command:

python -m benchmark.trustmem_bench --engine verimem   # 6/6, offline
python -m benchmark.trustmem_bench --engine mem0      # competitor (pip install mem0ai)

Same dataset, both engines (verimem · mem0, n=10 seed=42):

Axis	Verimem	mem0 OSS (raw)
Fabrication under absence	10/10	0/10
Destructive-update	10/10	10/10 · trivial (never reconciles)
Temporal integrity (as-of)	10/10	n/a — cloud-API-gated
Forget integrity (GDPR)	10/10	0/10 · delete leaves it in history(), verified live
Provenance honesty	10/10	0/10
Sycophancy resistance	10/10	n/a — cloud-API-gated
API coverage	6/6	40/60 (0.67)

Honest by construction (design §2): 100% on axes we build only proves no regression — the value is the competitor column and the two LLM-judged axes (still to come) where nobody scores 100%. mem0 runs in raw-store mode (their LLM extraction is out of scope, HaluMem measures that); maintainers are invited to PR an official run. Design + full reads: TRUSTMEM_BENCH_DESIGN.md.

Retrieval (LongMemEval-s, judge-free, fully local)

On the standard LongMemEval-s benchmark (full 500 questions, judge-free, fully local — result file): recall@5 0.8745 · hit@5 0.944 · MRR 0.856 (multilingual-e5-base, 768-d, fusion ON). On the Italian embedding A/B (result file), the e5 flip lifts Italian MRR 0.466 → 0.710 (+52%) at zero English regression (English recall@5 ties within noise: 0.80 e5 vs 0.82 legacy-MiniLM, 50-question slice).

End-to-end Italian smoke (benchmark/qa_italiano_smoke.py, the full recipe on an all-Italian store with accents): 13/13 factual + 6/6 abstention on trap questions = 1.000, dated transition resolved ("prima di trasferirsi → Milano"). Honest: a small hand-built synthetic set (easy, distinct facts) — a proof that recall AND anti-fabrication work in Italian end-to-end, not an external benchmark. The abstention 6/6 shows the anti-confab gate is not English-only.

New (2026-06-10): same-embedder comparative + full method notes in BENCHMARKS.md — Verimem vs a bare cosine baseline with the identical e5 model (the honest reading both ways: on pure ranking over a clean haystack the layer adds ~nothing — its value is provenance gating, crash-durable writes, per-model isolation and multi-tenancy, which a bare matrix doesn't have), plus the regime-dependence of the 2-stage cross-encoder rerank: default-ON on short facts (twice McNemar-validated, R@1 0.52→0.81 HARD / 0.53→0.68 fair-paraphrase) and auto-skipped on documents beyond the CE's 512-token window (ENGRAM_RECALL_RERANK=0 to opt out, ENGRAM_RERANK_MAX_DOC_CHARS to tune the guard, ENGRAM_RERANK_TOPN for the latency/quality knob).

Honest caveat: those are retrieval recall@k, not end-to-end QA-accuracy, so they are not directly 1:1 with the LOCOMO / LongMemEval QA scores Mem0 / Zep headline. The default fact-recall is single-hop cosine. The entity-graph + Personalized-PageRank engine (entity_kg.py, hippo_ppr_retrieve) is populated from the live corpus as of 2026-06-10: a deterministic zero-API extractor (entity_extract_lite.py) backfilled 7 570 entities / 75 959 co-occurrence edges / 22 609 fact-links from 2 181 facts, and PPR from real entities returns real facts (e.g. probe "Verimem" → 1 039 facts). Two limits, no spin: extraction is regex-tier (paths, modules, identifiers, proper nouns — not semantic OpenIE), and it runs as an idempotent backfill script (scripts/backfill_entity_graph.py), not yet wired into the live save path — new facts need a re-run to enter the graph.

System	Approach	Maturity
Mem0	flat vector + LLM summary; ships an MCP server (OpenMemory)	established, widely adopted
Zep / Graphiti	temporal knowledge graph, time-anchoring	commercial, mature
HippoRAG	OpenIE + Personalized-PageRank (the same hippocampal metaphor)	research; LLM-based OpenIE extraction (Verimem's PPR is live on real data too, but with a lighter regex-tier extractor)
Cognee	multi-step pipeline + pluggable ontology	established
Letta (MemGPT)	memory-as-OS, tiered paging	established
Verimem	sleep/Dream consolidation + self_model + critic-orchestrator gate + MCP-native + Italian-first + subscription-only	brand-new public release — 0 adoption yet

(Competitor benchmark scores and adoption numbers vary by source and drift over time — see our 2026-05 landscape survey for the sourced detail rather than trusting a number in this table.)

Where Verimem is behind — no spin: zero adoption (this is a fresh public repo); entity extraction is regex-tier (deterministic, zero-API) rather than semantic OpenIE, so common-noun entities are missed; and extraction is a backfill script, not yet wired into the live save path (HippoRAG / Zep extract at write time).

Where Verimem is genuinely different — the combination: (a) MCP-native integration (Mem0 also ships one via OpenMemory; the others are mostly SDKs you wire up yourself), (b) explicit sleep / Dream consolidation with a persistent self_model of the agent, (c) an adversarial critic gate on every write, and (d) running entirely on your Claude subscription — no API key, no per-token billing. No other system here combines all four. The entity-KG + PPR engine is live on real data (7 570 entities backfilled from the corpus, PPR returns real facts) — with the declared limits above (regex-tier extractor, backfill not yet in the write path).

Multi-tenancy (Mem0-parity, 2026-06-08): scope facts to user_id / agent_id / run_id with strict per-tenant isolation — a query for one tenant never sees another's — plus include_shared opt-in for global facts and scoped delete (delete_all(scope), dry-run default + per-fact undo). Zero schema change (the scope is a canonical user:…/agent:…/run:…/ topic prefix), so it works identically on both surfaces:

# MCP:  hippo_remember(proposition=…, topic="prefs", user_id="alice")
#       hippo_facts_recall(query=…, user_id="alice")          # bob's facts invisible
#       hippo_forget_scope(user_id="alice", dry_run=true)     # mem0 delete_all(scope)
# CLI:  engram facts add  -p "…" -t prefs --user-id alice
#       engram facts recall "…" --user-id alice [--include-shared]

Honest scope: this is the fact surface (Mem0's add / search / get_all / delete by scope); episode-level scoping is a declared follow-up, not yet shipped.

🔌 Verimem as MCP server — what Claude Code actually gets

When you wire Verimem into Claude Code (or any MCP host: Cursor, opencode, Cline, Continue, Zed), the host LLM sees 231 tools at session start. They cover the full memory-layer surface; below is the honest tour by category. Full per-tool reference: docs/MCP_QUICKSTART.md.

Category	Tools	What the host does with them
Recall	3	hippo_recall at task start → 5 past episodes (~3k tok). Find similar work, prime the answer.
Remember / Forget	3	hippo_remember to persist a fact (proposition + topic + confidence). Pure SQLite write, zero LLM.
Episodes	15	hippo_record_episode after every significant task. Plus pin/unpin, replay, classify, diff, clusters, dedup.
Facts / Semantic	15	hippo_facts_search (substring), hippo_facts_recall (cosine), hippo_facts_disagreement (conflict flag), merge/cluster/export.
Skills	50	Lifecycle (promote/retire/archive/recover/edit), discovery (top/similar/for/search), composition (lineage/cooccurrence/bundles), health (failure_audit/bottlenecks), import/export.
Hippo Dreams ⭐	7	hippo_dream_create_shadow → hippo_dream_propose → host LLM iterates submit_result per task → hippo_dream_diff → hippo_dream_adopt (atomic, rollback-safe). Zero internal LLM call.
Reasoning chains	12	hippo_reason, hippo_chain_validate, hippo_forward_chain, hippo_plan_strips, hippo_compose_macro, hippo_promote_chain.
Lineage / causal	8	hippo_causal_extract, hippo_skill_lineage_*, hippo_find_analogues, hippo_export_graph.
Audit / metrics	7	JSONL audit tail, metrics export, velocity stats.
Health / introspection	8	hippo_health, hippo_status, hippo_dashboard_overview, hippo_corpus_*.
Analytics / prediction	12	Outcome patterns, drift detection, RCA on failures, freshness/trust ranking.
Misc	~35	Rollup, prune, cross-agent consensus, briefings, provider switch.

Typical Claude Code session (real workflow, not synthetic):

User: "Refactor auth.py to use bcrypt."

Claude Code internally:

1. hippo_recall("refactor auth bcrypt") → finds 3 episodes from last month's sha256→bcrypt migration. Pulls them in (~3k tok).
2. hippo_skills_for("auth library swap") → retrieves a consolidated skill ("audit usage sites → introduce shim → flip default → remove old").
3. Edits auth.py applying the skill.
4. After tests pass: hippo_record_episode(...) + hippo_remember(proposition="auth.py now uses bcrypt cost=12", topic="project/X/auth").

Next month, "rotate bcrypt cost factor" → recall finds today's session, primes the answer instantly. Token bill drops to ~0 on the second-time-around work.

Context cost (honest): the ~231-tool registry is auto-loaded at session start → ~27k tokens of JSON schemas (~13% of a 200k window, ~3% of a 1M window). Skills/facts/episodes themselves do NOT auto-load — they're SQLite on disk, fetched per-call. A typical hippo_recall(k=5) returns ~3-5k tokens.

Subscription-first mode (ENGRAM_HOSTED=1): every tool that would call an LLM internally instead returns a structured payload the host can act on. Zero extra API spend — the host's Pro/Max plan does all the work. The Hippo Dreams pipeline runs entirely this way: propose (no LLM) → host iterates submit_result (host LLM call) → adopt (no LLM).

📈 Learning curve — what changes when memory is on (real LLMs)

5 iterations of the same task family on Anthropic Claude Opus 4.7, 8 tasks per iter (digit-sum suite). After the 3rd iteration Verimem has compiled the procedure into a deterministic macro that bypasses the LLM entirely:

Iter	Verimem tokens	Verimem latency	Raw LLM tokens	Raw LLM latency
0 (cold)	4225	4.47s	59	0.67s
1	1711 (-60%)	0.92s (-79%)	59	0.71s
2	687 (-84%)	0.52s (-88%)	59	0.85s
3	0 ✅	0.22s (-95%)	59	1.08s
4	0 ✅	0.24s (-95%)	59	0.69s

Break-even at iter 3. Macro fast-path hit rate: 70%. The agent stops asking the LLM for tasks it has solved before — your token bill drops to zero on recurrent work and your replies come back in 200 ms instead of 700 ms.

Bench data: data/bench_learning_curve_anthropic_n5.{results,by_iter}.json.

🎯 Held-out generalization — practical text tasks

Beyond digit-sum: 5 TRAIN tasks (URL parsing, date format, capitalize, reverse, word count) → sleep consolidate → 5 HELD-OUT tasks with fresh inputs of the same families.

Phase	Success	Rate
TRAIN	5/5	100%
HELD-OUT	5/5	100%

The agent retrieves 3 relevant skills per held-out task and applies them without per-task re-discovery. Bench data: data/bench_held_out_practical.{results.json,summary.md}.

📊 Headline — compositional generalization (real LLMs, 96 calls)

12 tasks at increasing skill-chaining depth (apply ROT3 + REVERSE in 1-5 chained transformations to fresh inputs). 4 providers, 1 iter:

Provider	Lv1	Lv2	Lv3	Lv4	Lv5	overall
Anthropic raw	100%	50%	0%	0%	0%	42%
Anthropic Verimem	100%	100%	100%	100%	100%	100%
DeepSeek raw	50%	0%	0%	0%	0%	25%
DeepSeek Verimem	100%	100%	100%	100%	100%	100%
OpenRouter raw	100%	0%	0%	0%	0%	33%
OpenRouter Verimem	100%	100%	100%	100%	100%	100%
Groq raw	50%	0%	0%	0%	0%	25%
Groq Verimem	100%	50%	100%	50%	100%	83%

The accuracy gap GROWS with composition depth. Single-shot LLMs collapse to 0% at Lv3; persistent-memory agents stay at 100% up to Lv5 on 3/4 providers. Empirical signature of compositional reasoning over consolidated skills.

Compiled-macro hit rate (skills bypassing the LLM entirely): OpenRouter 92%, Groq 75%, DeepSeek 58%, Anthropic 0% (Anthropic uses the full ReAct loop — the 100% accuracy comes from memory recall, not caching).

Bench data (replicable): data/bench_compositional_4providers.*.json.

python scripts/bench_with_without_hippo.py \
    --suite compositional \
    --providers anthropic,deepseek,openrouter,groq \
    --conditions raw,hippo_warm

Verimem — the product. The engine keeps the neuroscientific term engram — the physical substrate of a memory trace (Semon 1904; Lashley; Tulving; Tonegawa 2014 Nobel) — because that is what the system produces: inspectable, fitness-tracked, mergeable memory artifacts. Verified before they are stored: that is the Verimem part.

📚 Platform reference (start here)

If you want the end-to-end map of the moving parts — components, configuration knobs, task flow, MCP integration, multi-model bench harness, test isolation — read docs/PLATFORM.md. That doc is the contract; everything below is narrative.

For a 5-minute MCP integration walkthrough (Claude Code, Cursor, opencode, Cline, Continue) see docs/MCP_QUICKSTART.md.

🧬 What's new in cycle 213-250 (LLM-free emergent skill discovery, 2026-05-23)

The latest burst shipped a complete algorithmic skill discovery pipeline that finds, drafts, persists, and registers candidate skills FROM THE FACT GRAPH ITSELF without spending any LLM tokens on discovery + draft. The downstream LLM call is now an OPTIONAL polish step — not a discovery step.

detect_emerging_skills (213)         Louvain + topic purity + cohesion
  → normalize_topic (214/215)        family-key collapse + aggressive truncate
    → skill_drafter (217)            deterministic Markdown body + ranked keywords
      → skill_draft_persist (222)    ~/.engram/skill_drafts/<ts>/<name>.md
      → emerging_skill_register (229) topic=emerging_skill/auto-discovered/<name>
      → 4th Auto-Dream seed (219)    + adaptive_threshold (248/249) curve
      → promote_emerging_to_skill (235) gateway to SkillLibrary candidate

5 new MCP tools: hippo_emerging_skills_draft (218), hippo_skill_drafts_list (227), hippo_emerging_skills_register (232), hippo_emerging_skill_promote (236), hippo_emergence_pipeline_status (239).

6 helper scripts: dashboard, threshold_sweep, inspect_cluster, pilot_snapshot, bench_emerging_pipeline, plus the corpus-size-aware adaptive_threshold module (cycle 248).

4 empirical singolarità documented (A1 onesti, not marketing):

• #18 SELF-APPLYING LOOP — Auto-Dream discovery runs without any LLM token.
• #19 LINEAGE BACKWARD NAVIGATION — clp chain show <emerging_id> walks back 23 hops to source cluster ancestry.
• #20 SHADOW ZONE — at purity=0.2 the matrix surfaces 4 candidates (vs 1 at default 0.4): master-fact, antigravity-reverse, deep-clp, loop29-lineage.
• #21 OBSERVER-SHIFTS-EMERGENCE — re-running the same threshold sweep 4 min after registering shadow candidates moved 3 of them BACK under threshold. The session's own writes shift the Louvain partitioning (Heisenberg-like effect). Adaptive threshold curve (cycle 248-249) is the tuning scaffold; real cure (second-pass community detection) deferred.

A4 caveats:

• On the current 1921-fact corpus the pipeline surfaces 0 candidates at adaptive threshold 0.2 — requires p≤0.1 due to corpus fragmentation. Cycle 248-249 is FUTURE-PROOFING when corpus crosses 3000-4000 facts.
• Promotion rate H1 baseline 2.15% (vs cycle-174 audit 4.3%; retired pool grew). H1 multi-day pilot ready (cycle 244 scaffold).
• Cycle 228 H8c parallel_drafter FALSIFIED 1.28× < 1.5× target. Shipped as value-as-boundary.

See docs/emergence/README.md for quickstart and STATE.md for the full empirical snapshot.

🚀 What's new in v0.2.0 (production hardening)

The v0.2 push consolidates the prototype into a vendible system. See CHANGELOG.md and the four R&D reports (FINAL_REVIEW.md, BENCH_VALIDATION.md, QA_AUDIT.md, RND_MEMORIE.md, RND_UX.md) for the full picture.

• 13 CVE closed — RCE in /api/ide/run and the WS terminal (auth token + binary allowlist + shlex.split), SSRF blocklist in web_fetch, sensitive-path deny-list, plaintext API-key leak in /api/settings/providers redacted, computer-use kill-switch hotkey deny-list, dashboard CORS lock + session token, editfmt config-file deny-list, MCP server schema validation + audit log + rate limit, prompt-injection wrapper around external content (<untrusted_content>), Docker sandboxed Python executor (opt-in via HIPPO_PYTHON_EXEC_BACKEND=docker).
• Active memory v2 — 7 enhancements to the original 6 mechanisms + 5 new ones (11 total). All five new mechanisms are zero LLM call (pure numpy / string ops on existing memory). See RND_EXPLORATION.md for the diary of how they were found:
  • 7 — Working Memory Pruning (RND_MEMORIE.md): wake-loop char-budget compressor; critical for small-context models. Validated on Ollama qwen2.5:7b: −54 % token usage with lower variance vs unpatched build.
  • 8 — Trace Alignment / Reverse Replay (RND_TRACE_ALIGNMENT.md): Needleman-Wunsch on observation embeddings finds the exact divergence step between a failed run and its success-twin. Two-mode: action-divergence (same situation, different decision) + input-divergence (same tool, wrong file/query). Inspired by sharp-wave reverse replay in CA1 place cells.
  • 9 — Lateral Inhibition (Anti-Hebbian): when a winner skill consolidates on a task, its near-clone rivals are nudged away from that task vector. Földiák 1990 competitive specialisation. Empirically: −0.067 cosine differentiation at step 50 vs Hebbian-only baseline. Disabled by default; opt in via lateral_inhibition_enabled.
  • 10 — Spontaneous Reactivation: a default-mode rehearsal stage during sleep. Skills not used in N days get their last_used_at pushed forward by half the decay cutoff so they don't fall over the retirement cliff. Born & Wilhelm 2012 spaced-repetition substrate. Disabled by default; opt in via spontaneous_reactivation_enabled.
  • 11 — Salience by Surprise: replay priority gains a fourth term that boosts episodes whose num_steps deviates from the skill's average — Buzsáki 2015 prediction-error replay. Combined with multi-skill smallest-deviation logic so a typical-for-skill-A / anomalous-for-skill-B trace doesn't double-count. Disabled by default (sleep_replay_priority_surprise=0.0).
• Performance — hot paths are optimised: LRU-cached embeddings, vectorised skill clustering (corpus @ corpus.T), in-memory recall index with optional FAISS, mtime+size repomap cache. (Per-path speedups are workload-dependent — see tests/perf/ for the perf-test suite; the older "16×–4700×" figure had no reproducible benchmark behind it and was removed.)
• Architecture — dashboard.py 2 338 LOC monolith → 159 LOC entry-point + 11-file dashboard_routes/ package. LLM provider registry moved to providers.yaml + Pydantic ProviderSpec. New pydantic-settings-based Settings v2 singleton. Lightweight Alembic-style migrations.
• UX — Production-grade design system at engram/static/dashboard.css (WCAG 2.1 AA verified contrasts, 4 px scale, 1.25 type ratio, light theme). /skills page redesigned with KPI grid + responsive card grid + filter pills + accessibility. CLI banner with contextual tips and grouped /help.
• CI/CD — 3 OS × 4 Python = 11 jobs, dedicated security workflow (pip-audit, safety, bandit, ruff S-rules) running weekly. Multi-stage Dockerfile (~500 MB), non-root user, HEALTHCHECK. pip extras ([headless], [mcp-only], [tui], [vision], [full], [dev]) — default install is now minimal sane.
• Tests — 113 → 1072+ (+849 %). Coverage 46 % → 59 %. Ruff: 33 → 0 errors. Recent additions (FORGIA #27–#89): tests/test_bench_harness.py, test_bench_compare.py, test_bench_summary_md.py, test_bench_cli.py, test_bench_recall_ablation.py, test_clean_bench_data.py, test_jsonutil.py, test_corruption_guards.py, test_data_dir_isolation.py, test_auto_fallback.py, test_config_env.py, test_makefile_help.py, test_wake_used_macro.py, test_sleep_report_n_llm_calls.py, test_real_provider_smoke.py, test_mcp_e2e_smoke.py. Original suites preserved: tests/security/ (path traversal, SSRF, secrets redaction, prompt injection, executor isolation, editfmt sensitive, pentest validation), tests/test_settings.py, test_settings_v2.py, test_provider_registry.py, test_migrations.py, test_mcp_server.py, test_mcp_server_security.py, test_dashboard_api.py, test_cli.py, tests/perf/test_perf.py (10 benchmarks), tests/test_rnd_active_memory.py, test_trace_alignment.py, test_lateral_inhibition.py.

What Verimem does, in one breath

You give the agent a task in plain language. It thinks, uses tools (Python sandbox, file I/O, shell, web fetch, screenshots, webcam, vision LLM, computer use), retrieves any relevant past skills it has consolidated, and answers. Every conversation is an episode in memory. Every few episodes, you trigger a sleep cycle that distills new procedural skills (NREM), recombines them creatively (REM), merges duplicates, and promotes/retires by Bayesian fitness. Tomorrow the agent is genuinely better at what you asked it yesterday — and you can read every lesson it learned.

Verified working today

Multi-model bench harness (FORGIA #27, 2026-05-09): same task suite, 3 conditions (raw / hippo_cold / hippo_warm), 3-4 real providers, fail-isolated.

Headline result — memory_recall suite (the discriminative one):

Provider	raw	hippo_cold	hippo_warm
anthropic	0.50	1.00	1.00 (latency −56 %)
deepseek	0.50	1.00	1.00
openrouter	0.50	1.00	1.00

The 50 % raw failure is the 3 query tasks ("What was the value I told you?") — with no shared context, the LLM has no place to retrieve from. Verimem's recall pipeline retrieves the seed episode and the query phase succeeds 100 % on every provider. +50 percentage-point accuracy uplift, three different LLMs.

Hardened result — hard_memory_recall (12 tasks: direct + paraphrased + synthesis):

Provider	raw	hippo_cold	hippo_warm
anthropic	0.50	1.00	1.00 (latency −51 %)
deepseek	0.50	0.92	0.92 (lost the synthesis)
openrouter	0.50	1.00	1.00

The headline holds: +42–50 pp uplift across paraphrased queries and multi-step synthesis. DeepSeek lost the multi-step task (retrieved both facts but failed the addition) — Verimem provides the memory, arithmetic composition is on the model.

Skill-compounding suite (8 digit-sum tasks): hippo_warm latency −41 % vs hippo_cold on anthropic (compiled-macro fast-path engaging). Default trivia suite: raw wins (~50 tokens vs ~3 000), hippo costs structural overhead but proves end-to-end transport works.

All raw data committed at data/bench_*.{results,summary}.json. Full analysis in docs/PLATFORM.md.

Reproduce locally:

make bench-help          # list available task suites
make bench-mock          # offline smoke (no API key needed)
make bench-real          # run on every provider with an env key set
make bench-memory        # the discriminative recall suite
make bench-summary       # render the latest summary as markdown
make bench-csv           # CSV (Excel-friendly)
make bench-quick         # mock + 2 tasks (CI smoke)
make bench-clean         # dry-run of transient bench dirs
make stats               # project size + test count
make bench-compare BEFORE=... AFTER=...   # diff two bench summaries

The bench script supports many flags useful in CI: --quiet --max-tasks N --task-id ID --providers auto|csv --suite NAME --n-iter N --consolidate-every K --save-md --memory-stats --show-failures --print-config --list-providers --clean-data --output-dir PATH. See python scripts/bench_with_without_hippo.py --help.

Tool-use verified live across 4 providers (same task: write a real file to Desktop), all via native tool-use API (no fragile JSON-in-text parsing):

Provider	Model	Steps	Tokens	Outcome
Ollama (local, free)	qwen2.5:1.5b	2	4,655	✓ wrote file to disk
Ollama (local, free)	qwen2.5:7b-instruct	2	4,699	✓ wrote file to disk
Groq (free tier)	llama-3.3-70b-versatile	4	11,794	✓ wrote file to disk
Anthropic	claude-haiku-4-5	2	7,200	✓ wrote file to disk

Computer-use end-to-end verified live:

• ✓ shell_run: whoami, systeminfo, wmic, ver — Claude riassume sistema
• ✓ Task "apri Calc + screenshot + descrivi + chiudi" — 4 step, success
• ✓ vision_describe — Claude descrive logo NEXUS correttamente
• ✓ web_fetch + web_search (DuckDuckGo) — paper Nature/ScienceDirect
• ✓ desktop_screenshot 2560×1600 + describe via Anthropic vision
• ✓ Sleep cycle: 18 episodi → 6 NREM + 2 REM + 2 merge + 6 facts in 103s

Plan mode + auto-fallback chain verified.

Verimem is a working prototype of an idea: an LLM agent that becomes measurably more competent over time without ever updating its weights. It does so by mimicking the two-stage memory consolidation model from neuroscience (NREM slow-wave + REM paradoxical sleep). Every "lesson" the agent learns is a structured, versioned, fitness-tracked artifact you can read, edit, share, retire — not an opaque parameter shift.

Why this exists

Today's LLM agents have two ways to "learn":

1. Fine-tuning — costly, centralized, opaque, irreversible.
2. RAG / context — they "know" things but don't remember the session; no consolidation, no transfer, no growth curve.

The space in between — what humans actually do during sleep — is empty in production systems. Voyager (Wang et al.) introduced a skill library for Minecraft. MemGPT/Letta layered tiered memory. Reflexion added critique. But nobody has closed the loop with a consolidation cycle that:

• replays episodes (success and failure),
• extracts invariant patterns into procedural skills,
• recombines existing skills creatively,
• tests them under a fitness function,
• prunes the losers, promotes the winners,
• and measures itself with held-out tasks.

That loop is the bet of Verimem.

Architecture (one screen)

┌──────────────────────────────────────────────────────────────────────┐
│                            Verimem                                │
│                                                                      │
│  ╔═══════════════ WAKE ════════════════╗   ╔═════ SLEEP ══════════╗  │
│  ║  Task → Memory retrieval (skills    ║   ║  NREM: cluster        ║  │
│  ║          + similar episodes)        ║   ║   episodes →          ║  │
│  ║  ReAct loop with tool use:          ║   ║   distill skills      ║  │
│  ║   • run_python (sandboxed subproc)  ║   ║   + semantic facts    ║  │
│  ║   • syntax_check, find_function     ║   ║                       ║  │
│  ║   • submit_solution                 ║   ║  REM: pick 2 skills,  ║  │
│  ║  Reflexion-style self-critique on   ║   ║   propose hybrid      ║  │
│  ║   failure → 1-shot retry.           ║   ║                       ║  │
│  ║  Episode persisted to memory.       ║   ║  Curator: merge       ║  │
│  ╚══════════════════ ▼ ════════════════╝   ║   semantic dups       ║  │
│                      │                     ║                       ║  │
│              Episodes (SQLite +            ║  Pruning: Bayesian    ║  │
│              embeddings + causal graph)    ║   fitness → promote / ║  │
│                                            ║   retire skills.      ║  │
│              Skills (JSON files +          ╚══════════ ▲ ══════════╝  │
│              version chain + lineage DAG ──────────────┘              │
│              + Beta-Binomial fitness)                                │
│                                                                      │
│              Semantic facts (decoupled from time)                    │
└──────────────────────────────────────────────────────────────────────┘

Observability layer: every action emits a structured event →
   structlog → metrics registry → dashboard (FastAPI + vis-network).

What's actually inside

Module	What it does	Key novelty
episode.py	Episode + Trace — full ReAct trajectory, immutable	timestamped, embedding-indexed
memory.py	Episodic memory: SQLite + dense recall + causal graph (networkx)	clustering for replay; A→B causal edges via shared skill
semantic.py	Semantic memory: facts decoupled from time	Tulving-style separation
skill.py	Skill library — versioned JSON files + index	Bayesian Beta-Binomial fitness, lineage DAG, status lifecycle
tools.py	Sandboxed Python executor + AST analyzer	subprocess isolation, timeout, output cap
wake.py	ReAct loop with skill+episode injection	tolerant parser, Reflexion critique, A/B toggle
sleep.py	Multi-stage consolidation engine	NREM + REM + Curator + Pruning
prompts.py	All LLM prompts, in one auditable file	"experience as artifact" thesis
observability.py	EventBus + structlog + metrics registry	every step emits a typed event
dashboard.py	FastAPI + HTML dashboard	skill lineage graph (vis-network)
cli.py	typer + rich CLI	hippo run/wake/sleep/benchmark/skills/episodes/dashboard
benchmark/	18 HumanEval-style coding tasks + evaluator	wake/heldout split, Wilson CIs, two-prop z-test

Scientific anchors

• Two-stage consolidation: Walker & Stickgold (2004), Diekelmann & Born (2010). NREM consolidates declarative memory; REM enables creative recombination.
• Episodic vs semantic vs procedural: Tulving (1972, 1985).
• Fast hippocampal replay → slow cortical consolidation: McClelland, McNaughton & O'Reilly (1995) — the model Verimem's wake/sleep split mirrors.
• Reflexion: Shinn et al. (2023) — verbal RL on agents (no gradients) — used here for the self-critique retry.
• Voyager: Wang et al. (2023) — skill library for embodied agents — closest prior art for the procedural memory.
• Bayesian fitness for small N: Beta-Binomial conjugate prior — robust posterior mean even after 1–2 trials.
• Declarative → procedural transition: Anderson (1982) ACT-R, Logan (1988) instance theory — the basis for procedural compilation.
• Hippocampal forward sweeps: Pfeiffer & Foster (2013), Diba & Buzsáki (2007) — predictive replay before action — the basis for forward replay.
• Hebbian plasticity: Hebb (1949) "cells that fire together wire together" — the basis for the trigger-embedding drift on success.
• Counterfactual reasoning in episodic memory: Gershman & Daw (2017) — alternative trajectories during offline replay drive learning beyond mere reinforcement.

What makes the memory active (not just stored)

Most LLM-agent memory systems are passive: they retrieve past prompts and dump them into context. Verimem's memory is active — the act of using a skill makes it stronger, faster, and more discriminating. Six mechanisms compound:

1. Procedural compilation — once a skill has succeeded N times with high fitness, the DREAMER (during sleep) distills its successful traces into a parameterised macro: a list of tool calls with {{TASK}} and {{LAST_OBSERVATION}} placeholders. At wake time, when a strongly-matching task arrives, the macro is executed deterministically — zero LLM tokens, no model latency between steps. The skill is not just remembered, it is compiled the way deliberate actions become motor reflexes. (compilation.py)

2. Forward replay — before the wake loop fires, the agent looks up the top skill's past successful trajectories and projects an expected action sequence. The block is injected as ## PREDICTED PATH in the user prompt. This anchors the LLM's reasoning (less drift on familiar tasks) and lets us detect divergence (a learning signal). Pure retrieval — no extra LLM call. (wake.py:_forward_replay_block)

3. Hebbian skill embedding — every successful application drifts the skill's trigger embedding toward the task that just succeeded (new = (1 - α)·current + α·task, α = 0.05, then re-normalised). Skills become more retrievable for the kind of task they keep solving — the library shapes itself to its workload over time without any retraining. (skill.py:_hebbian_update)

4. Counterfactual REM — when a skill keeps failing (fitness < 0.5, trials ≥ 3), the dreamer doesn't just decrement Bayesian counts. It reads the failed trajectory and synthesises an alternative strategy — a candidate counterfactual skill with the failed skill as parent. The alternative competes for retrieval on future similar tasks; if it wins, it supersedes the broken approach without manual intervention. (sleep.py:_stage_counterfactual)

5. Schema formation — once a domain has accumulated enough skills (≥3 with cosine similarity ≥ 0.62 on triggers), the dreamer writes a SCHEMA: a meta-skill whose body is a one-line rubric for picking among the children. Lineage edges (relation='specialises') connect each schema to its specifics, building a 2-level hierarchy that becomes navigable as the library grows. Tulving's episodic→semantic transition. (sleep.py:_stage_schema)

6. Self-suggested practice — for skills sitting in the uncertain middle (fitness 0.45–0.65), the dreamer writes 2 concrete practice prompts that would plausibly trigger them. They appear in the dashboard's skill detail under "📚 Practice prompts" with one-click "▶ run in chat" buttons. Running a prompt feeds real evidence into the Bayesian fitness — so the skill is decisively promoted or retired instead of lingering in ambiguity. The agent literally suggests its own training set. (sleep.py:_stage_practice)

Together these turn memory from an archive into an organ that grows.

End-to-end demo (no API keys needed)

python scripts/demo_active_memory.py

Seeds a library with six skills, runs a complete sleep cycle (NREM + REM + Curator + compilation + counterfactual + schema + practice + pruning) using a scripted mock LLM, then re-runs a matching task to demonstrate the macro fast-path. Typical output:

Sleep cycle — six mechanisms fire in one pass
  duration    : 1.37s  (12 LLM calls, 3080 tokens)
  🔧 macros   : 1
  🌀 cf       : 1
  🌳 schemas  : 1
  📚 practice : 6 prompts written
  promoted    : 1   retired : 1

Wake — re-running a similar task uses the macro fast-path
  steps       : 2
  tokens used : 0
  llm calls   : 0  (macro fired, no model invoked)

Empirical evidence — macro speed-up

Run python scripts/bench_macro.py --repeats 3 --latency 0.5 to reproduce:

COLD (ReAct)    median  wall=1.296s  llm_calls=2  tokens=560
HOT  (macro)    median  wall=0.202s  llm_calls=0  tokens=0
→ speed-up      :   6.41x
→ time saved    : 1094 ms per task
→ token saved   :   560 per task

The simulated LLM models a real-world 0.5 s/call latency (typical for hosted models). On the hot path the macro fires deterministically — not a single token leaves the box.

Live dashboard

Visit /active-memory (or click "Active memory" in the nav) for a live KPI panel — compiled-macro coverage, Hebbian-tuned skills, counterfactual lineage, schema hierarchy. The /api/active-memory/stats JSON endpoint exposes the same metrics for external tooling.

Closing the loop — explicit user feedback

Every chat turn surfaces 👍 / 👎 buttons. Clicking them feeds the same Bayesian fitness machinery used by automatic outcomes — no separate "feedback model" or fine-tune. A 👎 on a successful turn flips the episode to failure and records a failure trial against each applied skill; mis-promoted skills decay back into the candidate pool. A 👍 simply boosts the trial count with a success.

So the library has three sources of evidence, all converging on the same fitness posterior:

1. automatic validator outcomes during the wake loop;
2. counterfactual REM evaluations during sleep;
3. explicit user up- / down-votes from the chat surface.

Install

Option A — Python (local)

git clone https://github.com/aureliocpr-ctrl/verimem.git
cd hippoagent
python -m venv .venv
source .venv/Scripts/activate           # or .venv/bin/activate on POSIX
pip install -e .
hippo dashboard                         # → http://127.0.0.1:8765

The first time you open the dashboard, you'll land on /welcome — pick a provider, paste an API key (or run Ollama locally — no key needed), test the connection, save. Everything is configurable from the ⚙ Settings page; no env-var voodoo required.

Option B — Docker

docker compose up
# or:  docker build -t hippoagent . && docker run -p 8765:8765 -v $PWD/data:/app/data hippoagent

The container picks up any *_API_KEY env vars you pass through. To use a host-running Ollama, the compose file already maps host.docker.internal.

Using Verimem

There are three ways to interact with the agent. They share the same backend (memory, skills, sleep cycles) — just pick what suits you.

1. Web dashboard

hippo dashboard → http://127.0.0.1:8765

page	what it's for
Chat ⭐	the main UI — type a task, get an answer with applied skills
Settings	pick LLM provider, paste API key, test connection, switch models live
Episodes	every task ever executed; click any row for the full ReAct trajectory
Skills	consolidated lessons with Bayesian fitness, status, lineage
Lineage	interactive graph of how skills derive from one another
Events	live event stream (every memory write, retrieval, LLM call)
Metrics	counters/histograms

2. CLI

hippo chat                       # interactive REPL: type tasks, /sleep, /skills, /quit
hippo run "your task here"       # one-shot
hippo wake --n-tasks 5           # run the bench wake-set
hippo sleep                      # consolidation cycle on demand
hippo benchmark                  # held-out evaluation
hippo skills list / show <id>
hippo episodes list / show <id>
hippo providers list / scan / models <p> / active

3. As an MCP server inside Claude Code / Cursor / opencode / Cline / Zed

Verimem ships an MCP (Model Context Protocol) server, so any MCP-aware client can use it as a memory layer. Add this to your client's mcp.json (or equivalent) — no API key needed: in hosted mode the host (Claude Code, Cursor, …) does all the LLM work, so memory tools like hippo_recall / hippo_remember cost zero extra API spend.

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "args": ["mcp"],
      "env": { "ENGRAM_HOSTED": "1" }
    }
  }
}

The engram and hippo commands are identical (same entry point) — use either. A provider API key is only required for the standalone agent (hippo_run_task, hippo chat), which runs its own LLM loop; the memory layer above never needs one.

Once registered, the host (Claude Code, Cursor, etc.) can call:

MCP tool	what it does
hippo_run_task	full wake loop — agent uses its own tools + skills
hippo_consolidate	trigger a sleep cycle
hippo_recall	semantic search over past episodes (embeddings)
hippo_recall_explain	semantic recall + per-component score breakdown
hippo_search	keyword search over episode task_text (LIKE)
hippo_episode_list	paginated listing of episodes (limit/offset/outcome)
hippo_episode_get	one episode in full (trajectory + critique)
hippo_episode_pin / hippo_episode_unpin	protect / release an episode from decay-pruning
hippo_forget	delete one episode by id (privacy / GDPR)
hippo_metrics_history	token-usage timeseries bucketed by day
hippo_skills_for	preview which skills Verimem would inject for a task
hippo_skill_promote / hippo_skill_retire / hippo_skill_edit	manual curation
hippo_skill_export / hippo_skill_import	portable JSON bundles (share skills between installations)
hippo_skill_test	render the prompt-context for a (skill, task) pair without calling the LLM
hippo_skill_top	top-k skills by fitness / recency / activity
hippo_skill_lineage	walk the parent_skills DAG ancestry of a skill
hippo_skill_compare	diff two skills (body / fitness / trials)
hippo_skill_similar	top-k skills by Jaccard overlap on body tokens
hippo_skill_describe	deterministic 1-line natural-language summary of a skill (no LLM)
hippo_skill_merge	manually merge skill A into B (sum trials, retire A, lineage tracked)
hippo_episodes_by_skill	every episode whose skills_used includes a given skill
hippo_provider_switch	switch the active LLM provider at runtime (anthropic / openai / groq / …)
hippo_remember	store one fact directly in semantic memory — no episode, no sleep cycle
hippo_facts_recall	semantic search over facts (cosine on proposition embedding)
hippo_facts_search	keyword/substring search over facts (LIKE on proposition)
hippo_facts_list	paginated listing of all facts (newest-first)
hippo_fact_forget	delete one fact by id (privacy / GDPR)
hippo_skills_search	keyword/substring search over skills (LIKE on name+trigger+body)
hippo_skill_bundles / hippo_compound_skills / hippo_skill_antagonists	structural introspection
hippo_status	counts + active provider
hippo_health	deep preflight at startup — 3-tier reachability + counts + flag + tool_count
hippo_stats	aggregate metrics (episodes by outcome, skills, token usage)
hippo_audit_tail	last N records of the MCP audit log (forensics)
hippo_prepare_task	HOSTED MODE: assemble prompt context (skills + recall) without calling any LLM
hippo_record_episode	HOSTED MODE: persist an episode the host LLM has just executed
hippo_consolidate_light	HOSTED MODE: dedup + promotion + retirement gates (no LLM)
hippo_plan_forward	REASONING / Pezzo B — hippocampal forward sweep (Pfeiffer & Foster 2013): beam-search top-k skill trajectories from start_skill. Pure local, no LLM.
hippo_plan_strips	REASONING / Pezzo A — STRIPS forward planner (Anderson ACT-R / Fikes & Nilsson): chain skills via preconditions/postconditions. BFS shortest-plan. Pure local.
hippo_find_analogues	REASONING / Pezzo C — structural analogy (Gentner 1983): find skills with high structural overlap but LOW semantic similarity. Surfaces transfer candidates the semantic recall misses.

And read MCP resources:

• hippo://skills/list and hippo://skills/{id} — consolidated skills
• hippo://episodes/recent and hippo://episodes/{id} — past trajectories

The result: your Claude Code (or whatever) gets a second-brain agent it can delegate to — and the second brain remembers across sessions, not just within one conversation.

4. Programmatically

from engram.agent import HippoAgent
agent = HippoAgent.build()
result = agent.run_task(
    task_id="my-task",
    task_text="Write a Python function that ...",
    validator=lambda ans: (bool(ans.strip()), "non-empty"),
)
print(result.episode.final_answer)
agent.consolidate()              # nightly sleep cycle

Desktop launchers (Windows)

powershell -ExecutionPolicy Bypass -File scripts\install_desktop_shortcut.ps1

This creates two shortcuts on your Desktop:

• Verimem Dashboard — double-click to launch the web UI (browser opens to http://127.0.0.1:8765).
• Verimem CLI — opens a shell with the venv activated, ready for hippo … commands.

Providers — bring any LLM you like

Verimem is provider-agnostic. Set ONE of these env vars (or run Ollama locally) and you're done. The first matching one wins (or force with HIPPO_LLM_PROVIDER=<name>).

family	provider (alias)	env var	base URL
native	anthropic	ANTHROPIC_API_KEY	Anthropic SDK
native	ollama	OLLAMA_HOST (defaults to http://localhost:11434)	local
US/EU	openai, openrouter, mistral, groq, xai (grok), perplexity, fireworks, together, cerebras, gemini (google), nvidia, huggingface (hf), deepinfra, hyperbolic, novita, lepton, anyscale, azure	<NAME>_API_KEY	each provider's /v1
China	moonshot (kimi), deepseek, qwen (dashscope/alibaba), zhipu (glm), baichuan, yi (lingyi/01ai), doubao (ark), hunyuan (tencent), stepfun (step), minimax, spark (iflytek)	<NAME>_API_KEY	each provider's /v1
local OpenAI-compat	lmstudio, vllm, localai, tabby	<NAME>_API_KEY (any non-empty)	localhost

Per-stage model overrides (Claude defaults are tuned; for other providers you usually want to set these):

export HIPPO_MODEL=qwen2.5:7b               # all stages
export HIPPO_MODEL_EXECUTOR=qwen2.5:7b      # ReAct loop only
export HIPPO_MODEL_DREAMER=qwen2.5:14b      # NREM/REM synthesis (smarter recommended)
export HIPPO_MODEL_CRITIC=qwen2.5:1.5b      # cheap critic

Discover what's actually reachable from your setup:

hippo providers list      # all known providers + env-var status
hippo providers scan      # query /v1/models on every configured one (real discovery)
hippo providers models kimi    # list one provider's models
hippo providers active    # which provider is selected right now

Examples:

HIPPO_LLM_PROVIDER=kimi     MOONSHOT_API_KEY=sk-...                              hippo wake
HIPPO_LLM_PROVIDER=deepseek DEEPSEEK_API_KEY=sk-... HIPPO_MODEL=deepseek-reasoner hippo wake
HIPPO_LLM_PROVIDER=ollama   OLLAMA_MODEL=qwen2.5:7b                              hippo wake
HIPPO_LLM_PROVIDER=groq     GROQ_API_KEY=gsk-...   HIPPO_MODEL=llama-3.3-70b-versatile hippo wake

For tests / dev:

pip install -e ".[dev]"

Quick run

# Tests (fully offline, mock LLM)
HIPPO_OFFLINE=1 pytest

# CLI
hippo --help
hippo run "Define a Python function that returns the n-th prime"
hippo wake --n-tasks 5             # run wake-set (records episodes)
hippo sleep                        # run consolidation cycle
hippo benchmark                    # run held-out tasks with consolidated skills
hippo skills list
hippo skills show <id>
hippo episodes list
hippo dashboard                    # → http://127.0.0.1:8765

End-to-end demo (the actual experiment)

python run_demo.py --n-wake 10 --n-heldout 8

What it does, in order:

1. Reset.
2. Baseline: held-out tasks, without skills, without past episodes — pure model.
3. Wipe the episodes the baseline accidentally created (clean slate).
4. Wake: run the wake-set; the agent records everything.
5. Sleep: NREM clusters episodes → distills skills; REM proposes hybrids; Curator merges semantic duplicates; pruning promotes/retires by fitness.
6. Hippo: re-run the held-out tasks with the consolidated skill library.
7. Compare: pass-rate, avg-steps, avg-tokens, skill-reuse-rate. 95% Wilson interval on rates. Two-proportion z-test for the gap.

The script saves a JSON report under data/reports/.

First-run observations

A small run (10 wake / 8 held-out) on Claude Haiku 4.5 produced these consolidated skills, derived autonomously from the agent's own behavior:

• Verify test cases before blaming algorithm (NREM, fitness=0.70)
• Validate test runner output before debug (REM hybrid, 0.70)
• Harden JSON code payloads end-to-end (REM hybrid, 0.70)
• Escape newlines in JSON code strings (the Curator merged 4 near-duplicates of this)

These are not generic "be helpful" stubs — they are concrete lessons the agent extracted from its own failure modes (mostly: malformed JSON in run_python calls when the code contained newlines).

The pass-rate gap on N=8 held-out is not statistically significant (p≈0.5 by two-prop z-test) — and that is the honest scientific finding at this scale. Ramping to N≥30 is the next experiment; the architecture is ready for it.

What makes this a prototype, not a toy

• ✅ 23 unit + integration tests passing
• ✅ Real LLM client + offline mock for deterministic CI
• ✅ Sandboxed Python execution (subprocess + timeout)
• ✅ Structured logging + event bus + metrics registry
• ✅ Bayesian fitness with conjugate prior (not naive ratios)
• ✅ Skill lineage DAG (networkx) — full provenance from episode → skill → REM hybrid → merge
• ✅ Web dashboard with skill lineage visualization
• ✅ Reproducible benchmark with deterministic seed split
• ✅ Statistical primitives (Wilson interval, two-prop z-test) for honest reporting
• ✅ A/B toggle (--no-skills) so baseline-vs-hippo is a single CLI flag

👁 Vision — works with any multimodal provider

The vision_describe tool dispatches to the right multimodal endpoint per provider. Defaults shipped (override via HIPPO_VISION_MODEL):

Provider	Default vision model	Status
Anthropic	claude-haiku-4-5-20251001	✓ verified live
OpenAI	gpt-4o-mini	✓
Google Gemini	gemini-1.5-flash	✓ free tier
Groq	meta-llama/llama-4-scout-17b-16e-instruct	✓ verified live, free tier
OpenRouter	anthropic/claude-haiku-4.5	✓ verified live
xAI Grok	grok-4	✓ (paid)
Mistral	pixtral-12b-latest	✓
Alibaba Qwen	qwen-vl-plus	✓
Zhipu GLM	glm-4v	✓
Moonshot Kimi	moonshot-v1-8k-vision-preview	✓
01.AI Yi	yi-vision	✓
ByteDance Doubao	doubao-vision-pro-32k	✓
NVIDIA NIM	meta/llama-3.2-90b-vision-instruct	✓
Together / Fireworks	Llama 3.2 Vision 90B	✓
HuggingFace router	Llama 3.2 Vision 90B	✓
Ollama (local)	llava (override with OLLAMA_VISION_MODEL)	✓ — needs ollama pull llava (or qwen2-vl, llama3.2-vision, bakllava, moondream)
DeepSeek	n/a	✗ DeepSeek V4 API doesn't accept image_url blocks

The dispatcher prefers HIPPO_VISION_MODEL env var > OLLAMA_VISION_MODEL (for Ollama) > the default in the table above. So you can mix: text inference on cheap-fast model, vision on a multimodal-capable one — same single call.

# Use Ollama for everything but route vision to Anthropic:
HIPPO_LLM_PROVIDER=ollama OLLAMA_MODEL=qwen2.5:7b \
  HIPPO_VISION_MODEL=ignored  # vision uses provider, set ANTHROPIC_API_KEY too
hippo dashboard

# Use DeepSeek for text, fallback to Groq for vision:
# (DeepSeek V4 has no vision; agent calls vision_describe → falls through to
# the configured fallback chain.)

🔁 Auto-fallback provider chain

When the active provider hits rate-limit / quota / 5xx, Verimem transparently tries the next configured provider. Configure in /settings → "Fallback chain":

primary: Anthropic Claude   (fast, paid)
   ↓ on 429/quota
fallback 1: Groq llama 70b  (free tier, very fast)
   ↓ on quota
fallback 2: Ollama qwen 7b  (local, always available)

Result: zero-downtime LLM access, scaling free → paid → premium automatically. Recoverable error patterns: 429, rate, quota, billing, credit, limit, 503, 504, timeout, connection, overload.

📋 Plan mode

Press 📋 Plan first in the chat instead of Send. The agent produces a numbered plan (3-7 steps) WITHOUT executing anything. Review it, then click ✓ Approve & execute to run the plan, or ✗ Reject.

Useful for: cautious computer-use tasks, multi-step shell ops, anything where you want to know what will happen before it does.

🔐 Permissions & Sandbox

Verimem exposes a single master switch plus 6 granular toggles in /settings:

Capability	Default	Description
Sandbox master	ON	When OFF, all permissions are unrestricted
Filesystem	home	strict (data/ only), home (user dir), full (anywhere)
Computer use	OFF	mouse/keyboard control via pyautogui
Webcam	OFF	capture frames + describe via vision LLM
Shell	OFF	arbitrary cmd.exe / /bin/sh commands
Web	ON	web_fetch + DuckDuckGo search
Vision	ON	describe images via multimodal LLM

Two preset buttons:

• 🔓 Unleash — sandbox OFF + all permissions ON. Full PC access.
• 🔒 Lockdown — sandbox ON, filesystem = strict, only web + vision.

When the sandbox is OFF the agent has full access to your machine: it can read/write any file, run any shell command, control your mouse and keyboard, capture your webcam, fetch the web, and describe images. Use this only with models you trust.

📱 Android / Termux

Verimem runs on Termux (Android 10+) with a few caveats:

pkg install python git ffmpeg-essentials
git clone https://github.com/aureliocpr-ctrl/verimem.git
cd hippoagent
python -m venv .venv && source .venv/bin/activate
# Skip pyautogui (no display server) and opencv (heavy):
pip install -e . --no-deps
pip install anthropic openai sentence-transformers numpy scipy scikit-learn \
            networkx pydantic structlog typer rich fastapi uvicorn jinja2 \
            python-dotenv httpx mcp textual pillow
hippo dashboard --host 0.0.0.0 --port 8765

Open http://<phone-ip>:8765 from any device on the same network. CLI works fully (hippo chat, hippo tui). Webcam/computer-use are no-op on Android, everything else (incl. Ollama via Termux+proot or remote, Groq, Gemini, Anthropic) works normally.

⚠ Security notes

• The Python sandbox is subprocess -I, not a real isolation layer. A task like "write a calculator and save it to the desktop" will actually write to your disk, because the LLM-generated code can call open(), os.makedirs, requests.get, etc. That's by design for this prototype (it's how you get an agent that does things), but it means only run Verimem against models and tasks you trust. For untrusted use, run the whole agent in Docker with no host volume mounts, or wrap the executor in seccomp/Firejail.
• API keys saved via the Settings page are written to data/user_settings.json in plaintext. Don't commit that file (it's in .gitignore). Production deployments should swap the storage for an OS keychain or Vault.

What's still toy

• Sandbox is subprocess -I, not seccomp/Firejail/Docker — fine for non-adversarial models, not for production.
• Benchmark is small (18 tasks). Real validation would need ≥100 tasks across more domains.
• No active learning: the agent doesn't choose which tasks to attempt to maximize learning.
• No explicit forgetting curve over time — only fitness-based pruning.
• Single-agent. The lineage DAG is ready for skill-sharing across instances; the marketplace is not built.

Repository layout

verimem/                       # repo root
├── engram/                    # core library (import engram)
│   ├── __init__.py
│   ├── config.py              # all hyper-params, single source
│   ├── observability.py       # event bus, logger, metrics
│   ├── llm.py                 # Anthropic client + MockLLM
│   ├── embedding.py           # sentence-transformers wrapper
│   ├── tools.py               # PythonExecutor, CodeAnalyzer, tool registry
│   ├── episode.py             # Episode + Trace
│   ├── memory.py              # EpisodicMemory (vector + causal graph)
│   ├── semantic.py            # SemanticMemory (facts)
│   ├── skill.py               # Skill + SkillLibrary (lineage, fitness, lifecycle)
│   ├── prompts.py             # all prompts, auditable
│   ├── wake.py                # ReAct loop with skill injection + critique
│   ├── sleep.py               # NREM + REM + Curator + Pruning
│   ├── agent.py               # high-level orchestrator
│   ├── cli.py                 # typer + rich CLI
│   └── dashboard.py           # FastAPI dashboard
├── benchmark/                 # task suite + evaluator + statistics
├── tests/                     # pytest suite
├── data/                      # episodes/skills/semantic/runs/reports (gitignored)
├── run_demo.py                # full baseline-vs-hippo experiment
├── pyproject.toml
└── README.md

License

MIT, but please cite the underlying neuroscience papers if you build on the consolidation idea.