Epimneme

Pronounced ep-im-NEE-mee. From Greek: ἐπί (epi-, "upon") + μνήμη (mnēmē, "memory") — meta-memory, a layer that sits upon memory itself.

Persistent memory for AI coding agents. PostgreSQL + pgvector backend, accessed via MCP or REST. Stop re-explaining your codebase every session.

---

Why Epimneme?

Every new chat with your coding agent starts from zero. You re-paste the same design decisions, re-explain the same gotchas, re-answer the same questions. Epimneme gives agents a long-term memory: facts, decisions, procedures, and a knowledge graph — all searchable, versioned, and deduplicated.

Agents connect via MCP (VS Code, Cursor, Claude Desktop, etc.) or through a plain REST API. Both use Bearer-token auth scoped per-project for multi-tenant safety.

Benchmarks

On LongMemEval (500 questions, 6 question types) with no LLM reranking and no benchmark-specific tuning:

Metric	Epimneme (no LLM)
Recall @ 1	79.0%
Recall @ 5	96.0%
Recall @ 10	98.8%
Recall @ 50	100%
NDCG @ 10	0.887
LLM cost	$0 / query

See benchmarks/BENCHMARK_RESULTS.md for the full write-up, including LoCoMo numbers and a per-category breakdown against MemPalace.

Quick Start

# 1. Clone
git clone https://github.com/alewman/epimneme.git
cd epimneme

# 2. Configure
cp .env.example .env
# Edit .env — at minimum set EPIMNEME_PG_PASSWORD to a strong value

# 3. Bring up Postgres + Epimneme
cp docker-compose.example.yml docker-compose.yml   # or edit in place
docker compose up -d --build

# 4. Wait for health, then create an admin API key
curl http://localhost:8000/health
docker exec epimneme python -m epimneme.manage create-key \
  --name admin --role admin

# Save the key it prints — it will not be shown again.

# 5. Create your first project
docker exec epimneme python -m epimneme.manage create-project \
  --name my-project --description "My awesome project"

Epimneme now listens on http://localhost:8000. See Connecting an Agent below.

Connecting an Agent

VS Code / Cursor / Claude Desktop (MCP over SSE)

{
  "mcpServers": {
    "epimneme": {
      "url": "http://localhost:8000/sse",
      "headers": {
        "Authorization": "Bearer epimneme_YOUR_KEY_HERE"
      }
    }
  }
}

Once connected, the agent sees these tools:

Tool	Purpose
session_start	Begin a session, receive previous context (summary, decisions, issues, entities)
session_end	Close session with summary + handoff notes for the next agent
remember	Store a memory (fact, decision, procedure, pattern, preference, observation, issue)
recall	Search memories by semantic similarity + keyword. deep=true for graph traversal
project_status	Get a project overview. Auto-registers new projects
entity_track	Add a node to the knowledge graph (file, module, concept, tool, person, library, config, command)
entity_relate	Add an edge (depends_on, part_of, uses, implements, …)
entity_explore	Traverse the graph from an entity (configurable depth + direction)

curl

curl -H "Authorization: Bearer epimneme_YOUR_KEY" \
  "http://localhost:8000/api/memories/search?query=database+config&project=my-project"

Full REST reference: docs/API.md (or run the server and visit /docs for the live OpenAPI UI).

What Gets Stored

Seven memory kinds. Pick the right one and recall works much better later.

Kind	Use for
fact	Discrete knowledge — "CLI entry point is main.py"
decision	Why something was done — "Chose PostgreSQL because pgvector + recursive CTEs cover both search and graph needs"
procedure	Step-by-step instructions
pattern	Recurring conventions — "All tests use pytest fixtures from conftest.py"
preference	Working style — "Prefers small PRs with single-purpose commits"
observation	General notes
issue	Known bugs, limitations, tech debt

Plus a knowledge graph of entities (files, modules, concepts) and relationships (depends_on, uses, part_of, implements, …) that recall can traverse when deep=true.

Features

• Hybrid retrieval — pgvector HNSW semantic search fused with PostgreSQL full-text + trigram keyword search via Reciprocal Rank Fusion.
• FSRS-inspired decay — memories fade without access, stabilize with repetition. The retrieval ranker boosts well-used memories.
• Dual dedup — SimHash (O(1) Hamming) catches minor rewordings; semantic cosine catches reworded but equivalent facts.
• Versioning — update_memory creates a new version instead of overwriting. Full history preserved.
• Conflict surfacing — when you store a new fact/decision similar to an old one, the response flags the potential conflict so the agent can resolve it.
• Periodic reflection — background job garbage-collects low-retrievability memories, consolidates clusters, and resolves detected conflicts. Pinned, persistent-project, decision, and procedure memories are exempt.
• Multi-tenant — projects are namespaces; API keys are project-scoped (or global admin).
• Dual auth — Bearer tokens for agents/programmatic access, OAuth passthrough (X-Forwarded-User) for browsers behind a reverse proxy.
• Rate limited — per-IP token bucket, honours X-Forwarded-For.
• Activity stream — in-memory ring buffer + text log for auditing.
• Dashboard — self-contained web UI for inspecting memories, entities, activity, and backups.

Architecture

┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│  VS Code /   │     │   n8n /      │     │  Browser     │
│  Cursor      │     │   Scripts    │     │  (OAuth)     │
│  (MCP/SSE)   │     │  (REST API)  │     │              │
└──────┬───────┘     └──────┬───────┘     └──────┬───────┘
       │  Bearer            │  Bearer            │  Reverse proxy OAuth
       └────────────┬───────┴────────────────────┘
                    │
          ┌─────────▼─────────┐
          │   FastAPI + MCP   │  ← /api/*, /sse, /messages, /, /health
          │   (port 8000)     │
          └─────────┬─────────┘
                    │
          ┌─────────▼─────────┐
          │   MemoryManager   │  ← sessions, memories, entities, decay, dedup
          └─────────┬─────────┘
                    │
          ┌─────────▼─────────┐
          │  PostgreSQL 16    │
          │  + pgvector       │
          │                   │
          │  vectors +        │
          │  full-text +      │
          │  graph (rCTE)     │
          └───────────────────┘

Full details: ARCHITECTURE.md.

Configuration

All settings are environment variables (EPIMNEME_*). See .env.example and the Configuration section of ARCHITECTURE.md for the full list.

Minimum required:

Variable	Notes
EPIMNEME_PG_PASSWORD	Must not be the literal string epimneme. The server refuses to start otherwise.

Deploying Behind a Reverse Proxy

The bundled docker-compose.example.yml has commented-out Traefik labels. Uncomment and adjust for your proxy. Recommended setup:

• Browser requests (/*) go through your OAuth/SSO middleware — users authenticate as admin.
• API / SSE requests (/api/*, /sse, /messages, /health) skip OAuth and use Bearer tokens instead. Do not apply gzip compression to /sse — it breaks SSE streaming.

Development

# Editable install with dev extras
pip install -e '.[dev]'

# Run tests
make test                # local (uses mocks, no DB needed)
make test-cov            # with coverage report

# Lint
make lint                # ruff check
make lint-fix            # ruff check --fix

See CONTRIBUTING.md for full contributor guidelines.

Security

If you discover a security vulnerability, please open a GitHub Security Advisory rather than a public issue. See SECURITY.md.

License

Apache License 2.0 — see LICENSE and NOTICE.

Credits

• Author & maintainer: Alewman
• Design & implementation assistance: Anthropic's Claude (via GitHub Copilot Chat and Claude Code). Large portions of the code, tests, migrations, reranking, and documentation were authored in close collaboration with Claude across many sessions — many of which Epimneme itself made possible by persisting the context.
• Built on FastAPI, PostgreSQL + pgvector, sentence-transformers, FlashRank, and the Model Context Protocol.
• Benchmarked against LongMemEval and LoCoMo; compared to MemPalace.