Skip to content

jigeagent/hermes-next

Repository files navigation

Hermes Next

A self-evolving memory system for AI agents. Python-native, zero-bridge, brain-like cognition.

Hermes Next fuses OpenViking vector storage with a Python-native MemOS cognitive engine, giving Hermes Agent agents persistent, self-evolving memory — no JSON-RPC bridge, no TypeScript dependency, no process overhead.

PyPI Python License Tests


Why Hermes Next?

Most agent memory systems today are TypeScript plugins bolted onto Python agents via JSON-RPC bridges — every memory operation crosses a process boundary, serializes and deserializes, and introduces a failure point.

Hermes Next is Python-native from day one. It runs in-process, shares memory with the agent, and eliminates the bridge entirely.

Feature Traditional (memos-local) Hermes Next
Language TypeScript + JSON-RPC bridge Python native
Search Brute-force cosine 6-step fusion (semantic + FTS5 + timeline + RRF + MMR + recency)
Lifecycle None (data grows forever) Auto-archive (90d) + confidence decay + pruning
Native memory sync ✅ MEMORY.md promotion + session_search fallback
Cross-agent sharing JSON-RPC ✅ OpenViking namespaces
Cognitive pipeline L1/L2/L3/Skill L1→Reward→L2→L3→Skill fully wired
Setup npm install + bridge config pip install

Bottom line: memos-local was built for the TypeScript era. Hermes Next is built for the Python-native era.


Quick Start

pip install hermes-next

Requires Python 3.10+ and a running OpenViking server (v0.3.22+).

Basic Usage

from hermes_next import HermesNextProvider

# Initialize
provider = HermesNextProvider()
provider.initialize(session_id="my-session")

# Every turn — automatic capture + retrieval
context = provider.prefetch("What did we discuss about RAG?")
provider.sync_turn(
    user_content="Tell me about RAG optimization",
    assistant_content="Key techniques: chunk size tuning, embedding selection...",
    session_id="my-session",
    tags=["rag", "optimization"],
)

# At session end — cognitive pipeline runs automatically
provider.on_session_end(messages=[])

# Check pipeline health
status = provider.handle_tool_call("memos_status", {})
print(status)

CLI

# Start the built-in viewer
hermes-next-viewer --port 8080

# Migrate from legacy memos-local-plugin database
hermes-next-migrate --old-db ~/.hermes/memos-plugin/data/memos.db

Architecture

                         Hermes Agent (Python)
                              │
                    HermesNextProvider
                     ┌────────┴────────┐
                     │                 │
              RetrievalPipeline   CognitivePipeline
              ┌─────┼───┬───┐    ┌───┼───┬───┬───┐
              │     │   │   │    │   │   │   │   │
           Semantic FTS5 Timel. MMR  L1  L2  L3  Skill
           (OV)   (SQLite)    (Rerank)  (Policy)(WM)(Cryst.)
                     │                 │
                  ┌──┴──┐          ┌───┴───┐
             OpenViking  SQLite  MEMORY.md  state.db
             (Vector)   (Cache)  (Native)  (Fallback)

4-Layer Retrieval Chain

  1. Hot Memory — MEMORY.md / USER.md (always in system prompt)
  2. Semantic Search — OpenViking vector store (primary, 95% of queries)
  3. FTS5 Full-text — Local SQLite cache (keyword fallback)
  4. session_search — Hermes Agent native state.db (last resort)

Cognitive Pipeline

Every session triggers an automatic induction chain:

L1 Traces → Reward Backprop → L2 Policy Induction → L3 World Model → Skill Crystallization
   ↓              ↓                  ↓                   ↓                 ↓
 Raw turns    Score each       Extract reusable       Build domain      Package as
              interaction     behavioral patterns     knowledge         invocable skills

Configuration

# ~/.hermes-next.yaml
openviking:
  base_url: "http://localhost:1933"

cognitive:
  enable_l2_induction: true
  enable_l3_world_model: false   # opt-in, GPT-intensive
  enable_skill_crystallization: false

lifecycle:
  trace_retention_days: 90
  policy_decay_rate: 0.03

integration:
  sync_memory_md: false           # promote to Hermes Agent MEMORY.md
  session_search_fallback: true    # fallback to native state.db FTS5

Tools Exposed to the Agent

Tool Description
memos_search(query, k) Semantic search across all memories
memos_get(trace_id) Read a specific trace
memos_timeline(limit) Recent memory activity
memos_status Pipeline health + promotion stats

Project Structure

hermes-next/
├── hermes_next/
│   ├── ov/            # OpenViking REST client
│   ├── memos/         # MemOS cognitive engine
│   ├── cache/         # SQLite local cache (FTS5 + vector)
│   ├── retrieval/     # 6-step fusion pipeline
│   ├── integration/   # Hermes Agent native memory bridge
│   └── viewer/        # Built-in SPA dashboard
├── tests/             # 141 tests, 0 flaky
└── docs/              # Architecture + upgrade guides

Comparison with memos-local-plugin

Aspect memos-local-plugin Hermes Next
Runtime TypeScript, separate process Python, in-process
Communication JSON-RPC over stdio Direct function calls
Vector search Brute-force cosine (SQLite) OpenViking (real vector DB)
Full-text search FTS5
Memory lifecycle Auto archive + decay
Native Hermes sync MEMORY.md + session_search
Cross-agent sharing Per-process OpenViking namespaces
Viewer HTTP + SSE HTTP SPA
Pipeline 3-tier retrieval 6-step fusion + cognitive

Roadmap

  • v0.4.0 — Feedback loop (user 👍/👎 → policy update → better retrieval)
  • v0.5.0 — Decision Repair (failure patterns → prevention)
  • v0.6.0 — Hub cross-agent search (native, no OV dependency)

See docs/strategic-positioning.md for the full strategy.


License

AGPL-3.0 — This project is a derivative of OpenViking (AGPL-3.0).


Built for agents that learn from every conversation. 🤛

About

Next-generation memory provider for Hermes Agent — fusing OpenViking vector storage with MemOS cognitive engine

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages