Local-first AI memory for agents. Zero third-party dependencies, SQLite on your own disk, MIT.
We build memory infrastructure that does not need a cloud account to work — and we build it using itself.
| Repository | What it is |
|---|---|
| mnemosyne | The engine. 52 MCP tools (41 native + 11 client-compatible), Python API, REST API, CLI. Local-first, zero dependencies. |
| mnemosyne-bench | Reproducible benchmarks for AI agent memory systems, with the scoring harness included. |
ai-memory.net — the same engine with a web console, API keys and subscriptions. Self-hosting stays MIT and unlimited; no feature is paywalled in the open-source build.
From the benchmark harness, re-runnable against your own corpus:
| Benchmark | Mnemosyne OS | Mem0 (published) |
|---|---|---|
| LongMemEval | 96.2 | 94.4 |
| LoCoMo | 94.8 | 92.5 |
| BEAM 1M | 68.5 | 64.1 |
| BEAM 10M | 53.9 | 48.6 |
Mem0's figures are the ones Mem0 publishes at mem0.ai/research.
- Issues: the relevant repository's issue tracker
- Everything else: ai-memory.net/#/contact