BEAM (Bilevel Episodic-Associative Memory): a working-memory tier (hot context, auto-injected before LLM calls, TTL eviction) consolidates into an episodic long-term tier, alongside a TripleStore temporal knowledge graph with version chains and as-of queries. Optional LLM-driven fact extraction and entity extraction at write time; a `sleep` command runs consolidation. Memory banks give per-domain isolation. Ships Hermes-first (native plugin with lifecycle hooks) but is framework-agnostic via MCP, Python SDK, and CLI.
- Storage
- A single SQLite file — sqlite-vec for dense vectors plus FTS5 for keyword search, no external services. MIB information-theoretic binarization compresses 384-dim float32 embeddings to 48-byte binary vectors (32x reduction) with Hamming distance computed inside SQLite. Optional Mnemosyne Sync replicates deltas between instances (desktop↔VPS) with optional client-side XChaCha20-Poly1305 encryption so the sync server sees only metadata.
- Retrieval
- Hybrid scoring inside SQLite: 50% vector similarity + 30% FTS5 rank + 20% importance (weights configurable via env vars), with optional temporal recency boost (configurable half-life) and knowledge-graph as-of queries. Default embedder bge-small-en-v1.5; local embeddings via fastembed/sentence-transformers extras or any OpenAI-compatible remote embedding API.
- Self-host
- Self-host: trivial
- License
- MIT
- Pricing
- Free, MIT-licensed OSS; no hosted or paid tier — development is sponsor-supported (compute-credit sponsors listed at mnemosyne.site/partners). Sync server is self-hosted (Docker, bare metal, Fly.io). · Free / OSS
- GitHub stars
- 1,855
- Last release
- 2026-07-19
- Last commit
- 2026-07-25
- First catalogued
- 2026-07-25
Strengths
- Genuinely local-first, minimal-infra: one pip install, one SQLite file, no vector DB, no Postgres, no external services — core profile runs in ~50 MB RAM on a Raspberry Pi
- In-process reads with no HTTP round-trip; MIB binary vectors (32x compression) keep Hamming-distance search entirely inside SQLite
- Broad integration surface for its age: MCP server plus native adapters for Hermes, OpenWebUI, OpenClaw, and Pi, a Python SDK, a CLI, and a published integration template
- Temporal knowledge graph (version chains, as-of queries) and per-domain memory banks alongside vector + FTS5 hybrid recall
- Self-hosted multi-device sync with optional client-side XChaCha20-Poly1305 encryption — the sync server sees only metadata
Watch out
- All benchmark numbers are self-reported from its own README (LongMemEval 98.9% Recall@All@5 on a 100-instance subset; BEAM 65.2% at 100K) — and its own BEAM table shows Hindsight higher (73.4%) at the same scale
- 'Zero-dependency' applies to the core profile only, which has no local embeddings and needs an OpenAI-compatible remote embedding API; local embedding extras add ~800 MB–1.5 GB of dependencies
- Young (created 2026-04) and primarily single-author (AxDSan / Abdias J); PyPI project metadata still points at the author's personal repo rather than the mnemosyne-oss org
- Confusing dual release tracks in one repo (main package v3.x alongside a v0.x companion line); a write-approval-bypass + path-traversal advisory (GHSA-v825-68pf-4mfh) was patched 2026-07-19
- Hermes-first defaults leak through — e.g. the default data directory is ~/.hermes/mnemosyne/data even outside Hermes
Best for
- Local-first or offline personal agents where memory must stay on-device with no cloud in the read path
- Coding-agent memory over MCP (Claude Code, Cursor, Codex CLI, Windsurf) with sub-millisecond in-process recall
- Low-resource deployments (Raspberry Pi, 1 GB VPS) that can't carry a vector DB or Postgres
- Hermes Agent users wanting a native memory provider with automatic context injection
How it integrates
Benchmark results
| Benchmark | Value | Backbone | Trust | Source |
|---|---|---|---|---|
| longmemeval | 98.9 recall | — | Self-reported | Mnemosyne OSS ↗ |
| beam-100k | 65.2 accuracy | Llama 3.3 70B (via NVIDIA API) | Self-reported | Mnemosyne OSS ↗ |
| beam-100k | 20 recall | — | Self-reported | Mnemosyne OSS ↗ |
| beam-100k | 372 latency-ms | — | Self-reported | Mnemosyne OSS ↗ |
| beam-1m | 20 recall | — | Self-reported | Mnemosyne OSS ↗ |
| beam-1m | 493 latency-ms | — | Self-reported | Mnemosyne OSS ↗ |
| beam-10m | 20 recall | — | Self-reported | Mnemosyne OSS ↗ |
| beam-10m | 35 latency-ms | — | Self-reported | Mnemosyne OSS ↗ |
Sources
- Mnemosyne README — BEAM architecture, hybrid scoring weights, MIB binary vectors, platform table, sync + security model, MIT license (vendor)
- GitHub API — 1808 stars, MIT license, created 2026-04-05, pushed 2026-07-25, not archived, not a fork (third-party)
- PyPI — mnemosyne-memory v3.14.0, MIT; project URLs point at AxDSan/mnemosyne (author repo) (vendor)
- Integration guides — per-platform setup for Claude Code/Cursor/Codex/Windsurf MCP, OpenWebUI, OpenClaw, Pi, Hermes (vendor)
- v0.5.0 release (2026-07-19) — security fix for GHSA-v825-68pf-4mfh (write-approval bypass, path traversal); dual v3.x/v0.x release tracks (vendor)
Last verified 2026-07-25 · updated by discover-frameworks