Bionic memory for AI agents

Give your AI a brain that survives the next session.

Mnemosyne is a local-first memory system with knowledge graph retrieval, predictive validation, dream consolidation, and governed skill evolution through MCP.

v7.2.0 release published GitHub Pages enabled Discussions open
Star on GitHub Read the docs FAQ 
git clone https://github.com/nianpangzhi233/Mnemosyne-AI-Memory.git
cd Mnemosyne-AI-Memory
python setup.py

python scripts/graph_write.py --content "gzip bodies must be decompressed" --principle "check Content-Encoding first"
python scripts/graph_query.py --hybrid-search "request body parse failure" --layer L0
python scripts/graph_dream.py --full

Not a diary. Not just RAG.

Mnemosyne treats experience as living knowledge: it can be searched, contradicted, decayed, consolidated, and eventually crystallized into reusable skills.

GraphRAGKnowledge graph + vector search

Hybrid, precise, creative, vector, and keyword search across semantic, causal, temporal, and entity dimensions.

PredictionMemory that knows when it applies

Preconditions and predicted outcomes let memories warn the agent before repeating old mistakes.

DreamAutomatic consolidation

Fast/Slow dream cycles discover relations, decay weak memories, and keep the graph clean.

Skill MemorySkills with evidence

Experience clusters grow into skills only after live tests, feedback, and governance checks.

MCPAgent-native integration

Use memory_write, memory_search, memory_inject, and skill tools from any MCP-compatible AI client.

DashboardObservable by default

Inspect memory health, search results, graph shape, dream logs, and skill evidence in Streamlit.

Visible by design.

The dashboard and architecture view are part of the product story, not decoration. Memory systems need evidence people can inspect.

Mnemosyne architecture Mnemosyne dashboard preview

Skills evolve with gates.

Prompts are easy to write and easy to trust too much. Mnemosyne keeps a conservative evidence flow before default injection.

01

Experience cluster becomes an embryo.

02

Draft skill is generated and mirrored to disk.

03

Darwin-style live tests compare baseline vs with-skill.

04

Feedback tracks success, miss, trigger mismatch, and misleading output.

05

Only low-risk stable skills enter default injection.

Built for people who keep teaching an AI the same lesson and are tired of watching it forget.

Local first Try it in five minutes

Python 3.10+, SQLite, local embeddings, optional LLM review, REST API, MCP server, CLI, and dashboard. No external service is required for the core path.

Get started
What you'll find
This site is built to help visitors understand the project fast: what it does, why it matters, how to start, and where to read the public docs.

Need the quick answer?

These public docs are written for people arriving from search, social posts, or GitHub recommendations.