Agent memory,in one file.
Mnemosyne is a local-first memory engine for Hermes, Claude Code, Cursor, Codex and any MCP client. One SQLite file, one pip install, no server.
How a memory lives
First it remembers.
remember() writes to working memory with an importance score and an embedding, in about 12 ms.
>>> from mnemosyne import remember, recall, sleep>>> remember("Ana prefers dark mode", importance=0.9, scope="global")# stored in working memory, 12 ms
Try it without installing anything
A client-side simulation: keyword matching, no model, nothing leaves your browser. Tell it something, ask about it, put it to sleep.
A simulation. The real engine adds vector search and catches different phrasings.
Install the real thingWhat it does for you
Works with the network cable unplugged.
Every query is a local SQLite read. Zero network calls, ever.
Keeps every memory on your own disk.
One file you can copy, back up or delete. No telemetry.
Finds things you phrased differently.
Hybrid ranking: 50% vector, 30% full-text, 20% importance.
Tidies itself while the agent sleeps.
auto_sleep consolidates old working memories into episodic summaries.
Remembers when a fact stopped being true.
Temporal triples carry valid-from and valid-to dates.
How many memories will your agent keep?
49.1MBin one SQLite file
Roughly 2 KB per memory, embedding included: the text, its embedding vector and the row around it. An estimate from the schema, not a measurement.
The published numbers
Measured on CPU at 1,000 entries with sqlite-vec and FTS5, no GPU, as listed on the performance page of the docs.
- 12msWrite
- 45msHybrid search
- 15msVector search
- 5msFull-text search
BEAM benchmark, v3.0.0, May 2026
Against the published ICLR 2026 baselines. The end-to-end score is LLM-as-judge at 100K; the scale and abstention figures are pure retrieval.
- 65.2%end-to-end QA at 100K context
- 20%recall@10 held flat from 100K to 10M entries
- 100%abstention when the corpus has nothing
Three tiers in one file
Bilevel Episodic-Associative Memory: three SQLite tables in one file. Hot context on top, long-term below, a scratchpad for the agent's own reasoning.
Working memory
Hot, recent context, auto-injected into prompts. Session-scoped by default.
Episodic memory
Long-term storage with sqlite-vec and FTS5. Where sleep() puts what mattered.
Scratchpad
A temporary reasoning workspace, cleared per session.
Runs where your agent already runs
- Hermes
- Claude Code
- Cursor
- Codex
- OpenWebUI
- OpenClaw
- Windsurf
- MCP
Any MCP client works: point it at mnemosyne mcp and 30+ tools register. There is a Python SDK for everything else.
Against the cloud memory services
What you gain and what you trade. Prices and features as published in May 2026; tell us if one has changed.
| Feature | Mnemosyne | Honcho | Zep | Mem0 |
|---|---|---|---|---|
| Cost | Free forever | $$$ Paid (credits) | $$$ Paid (Flex+) | Freemium ($0-$249/mo) |
| Hosting | Local - your machine | Cloud only | Cloud / BYOC | Cloud only |
| Privacy | 100% local, zero exfil | External API calls | External API calls | External API calls |
| Offline mode | Yes - airplane mode | No | No | No |
| Setup | pip install | Docker + API keys | Docker + Postgres | API key + signup |
| Vector store | sqlite-vec (built-in) | pgvector (external) | pgvector (external) | pgvector (external) |
| Full-text search | FTS5 (built-in) | Separate service | Separate service | Separate service |
| Auth required | None | Supabase auth | OAuth / API key | API key |
| Rate limits | Unlimited | Plan-dependent | Credit-based | Plan-dependent |
| Data ownership | You own the SQLite file | Vendor-hosted | Vendor-hosted | Vendor-hosted |
| Export / import | One JSON file | Limited | Limited | Limited |
| Dependencies | Python stdlib + ONNX | Docker, Postgres | Docker, Postgres | pip + API key |
| Memory architecture | BEAM (3-tier) | Session + facts | Graph RAG + facts | Session + facts |
| Auto-consolidation | Sleep cycles built-in | Manual / paid | Manual | Manual |
| Temporal triples | Native with validity | No | No | No |
| BEAM-100K | 65.2% (100K, v3.0.0) | 63.0% (BEAM paper) | Not published | Not published |
The trade: you keep your own backup (one file), and there is no hosted team sharing. Mnemosyne is built for individual agents and single-machine deployments.
Give your agent a memory
One pip install. Free, MIT licensed, and the file is yours.
