HyperMnesia sits in a narrow slot: not chat memory — self-hosted knowledge for a coding
agent. A hand-authored map of file→component→rules, plus doc-RAG, plus bi-temporal personal
memory, all in your own Postgres. The neighbours below solve adjacent problems well; almost none
of them do the deterministic Tier 1 (path → component → must/should, no model in the loop).
Facts below were checked against each project's own repo/site (links inline). External projects change — the rows below were re-checked against primary sources on 2026-09-10, and self-reported benchmark claims are labelled as such.
| Project | Bet | Store | License | Not what HyperMnesia is |
|---|---|---|---|---|
| HyperMnesia | file rules before an edit + docs + personal memory | Postgres + pgvector | MIT | doesn't index code; map is hand-authored |
| hypermnesia-mcp / Cortex | persistent memory + self-curating wiki for coding agents | SQLite (default) or Postgres + pgvector | MIT | decay is central; no file→rule map |
| agentmemory | auto-capture coding-agent sessions | SQLite + iii-engine | Apache-2.0 | no deterministic component/constraint map |
| Vestige | "find the cause of a bug, not the lookalike" | SQLite + FTS5 + HNSW | AGPL-3.0 | FSRS decay, single 25MB binary |
| Hindsight | an agent that learns (retain/recall/reflect) | Postgres+pgvector / cloud | MIT | conversational/enterprise memory at its core |
| mem0 | memory API for any agent | library / self-host / cloud | Apache-2.0 | not about repo invariants |
| supermemory | "the memory API" | self-hosted single binary, or cloud | MIT | a general memory engine, not a repo constraint map |
| Basic Memory | memory as Markdown on disk | Markdown files + a local SQLite index (MCP) | AGPL-3.0 | files are the source of truth, not SQL |
| Claude memory | managed chat memory | Anthropic-hosted | proprietary | not your SQL, not a code map |
Notes on the neighbours worth stating plainly:
- agentmemory is the closest market neighbour: hooks, hybrid search, SessionStart injection, "don't re-explain." It deliberately does not index code either — it recommends pairing with external code-graph tools (e.g. codegraph, a separate third-party project it doesn't ship). That's the same split HyperMnesia makes with an LSP/Serena. Its headline efficiency claim is self-reported (~1.9k tokens/session per its README).
- Vestige is closest in philosophy ("this isn't RAG") but bets differently: novelty-gated writes and FSRS-6 decay. HyperMnesia rejects decay on purpose (see below).
- Hindsight is a heavier animal — fact/experience/observation/mental-model networks, RRF + a
cross-encoder, "knowledge pages" as a living wiki, and an
npxinstaller for coding agents. It claims state-of-the-art on the LongMemEval benchmark; that's a vendor claim, and supermemory claims #1 on the same benchmark, so read "SOTA" as contested. HyperMnesia doesn't play on that field — and shouldn't, if the niche is "control over a repo," not "SOTA chat memory." - hypermnesia-mcp / Cortex is the closest neighbour by feature list, and the reason the name collision below matters. Overlapping: nine Claude Code lifecycle hooks, "a self-curating per-project wiki", supersession with history ("Corrections supersede rather than overwrite. The new memory records what it replaces, the old one is demoted in recall"), weighted RRF over five signals — vector, full-text, trigram, heat, recency — then a cross-encoder rerank, and SQLite by default with PostgreSQL + pgvector optional. Diverging: decay is load-bearing there ("Memories carry heat that decays unless replay reinforces them"), which HyperMnesia refuses on purpose; and its README documents no deterministic file→component→rule map — searched for glob/constraint/ invariant wording, no hits. Its LongMemEval figures are self-reported (Recall@10 98.2% at v4.14.1, 97.8% at v4.20.0), though published with reproduction artefacts and a code SHA.
- Name collision — verified, not a rumour.
hypermnesia-mcpis on PyPI: version 4.21.0, MIT, 31 releases between 2026-06-19 and 2026-09-10, repository cdeust/Cortex. It is not a namesake in another niche — it is persistent memory for AI coding agents, the same words for the same kind of tool (see the bullet above). An earlier version of this note called the collision an unverified LinkedIn announcement by another author and said no matching public repo existed; that was wrong on both counts. What to do about the name is the owner's call; this document only records the facts.
Deterministic path → component → must/should plus a 1-hop graph. Elsewhere "rules" are a
recovered snippet or an LLM-recalled fact. Asked "what may not be imported into
src/api/routes.py?", HyperMnesia answers without a model; similarity-based stores retrieve
something that looks related. Beyond that: abstention instead of top-k noise, supersede-with-
history instead of overwrite, a review queue gating merges, fail-open, and an explicit refusal to
embed code. Not all of that is exclusive — hypermnesia-mcp/Cortex ships supersession-with-history
and a self-curating wiki too. The map lives in SQL — readable and editable, not trapped in a
model's head.
Being straight about the edges, because a deterministic map that has gone stale lies more confidently than search does:
- The map is hand-authored — that's the value and the liability. While the seed tracks the
tree, Tier 1 is excellent. When files move and globs don't, it silently stops resolving. This is
why
ci/freshness.pyexists (orphan-glob and stale-doc detection, plus an informational count of constraints with no source document) — run it in CI or on a schedule so decay is loud. - Ingest is markdown-only. Decisions that live in PRs, commits, or chat don't reach the map by themselves — someone has to write them into docs or the seed.
- No forgetting. Personal memory accumulates in
mem.memories(active_memoriesis a view over it, filtered to status and validity window); there's no decay curve, so quality rests on extraction + the review queue, not on eviction. Deliberate (an agent's "never do X here" shouldn't fade), but it means volume is managed by consolidation, not time. - Consolidation is pairwise cosine over active memories — simple, and O(n²) as memory grows. Fine for a personal/homelab store; a larger deployment would want blocking/ANN before this bites.
- Operationally heavier than
npx-a-single-binary neighbours: Postgres + an embedder (Ollama/TEI) + the Rust MCP server. Good for "own your store," not for "install and forget."
Nearest by feature overlap — and by name — is hypermnesia-mcp / Cortex; nearest by market position, agentmemory; by local-first spirit, Vestige; by memory-as-a-system, Hindsight. HyperMnesia's distinctive bet isn't RAG and isn't hooks — it's that a file's rules are data, not a retrieval result. While the map is alive that's an advantage; the day the globs fall behind, the others still return "something similar" and HyperMnesia would return a confident wrong answer — which is exactly why the freshness check and the hand on the seed matter.