로컬 LLM을 위한 영구 기억 계층, Mnemo

mnemo Local-first AI memory layer for any LLM. Persistent knowledge graph, entity extraction, semantic retrieval — no cloud required. What is mnemo? Most LLMs forget everything the moment a conversation ends. mnemo fixes that. mnemo is a sidecar service that watches every conversation you feed it, extracts named entities and relationships using an LLM, builds a persistent knowledge graph in SQLite, and injects relevant context back into future prompts — automatically, in under 50ms. It works with Ollama (fully local, free), OpenAI, Anthropic, or any OpenAI-compatible API. It ships as a single static binary with zero cloud dependency. How it works your app │ ▼ POST /ingest ──► entity extraction (LLM) ──► knowledge graph (SQLite + petgraph) │ POST /retrieve ◄── scoring + ranking ◄── graph traversal + full-text search │ ▼ context_prompt ──► inject into your LLM prompt You POST raw text to /ingest (a conversation turn, a document, a note). mnemo sends it to your configured LLM and extracts entities (people, tools, places, concepts) and the relationships between them. Entities are deduplicated by name+type, aliases are merged, and everything is written to SQLite. The in-memory petgraph is updated atomically. On POST /retrieve , mnemo runs a 6-stage pipeline: full-text chunk search → entity name search → graph expansion (BFS over the knowledge graph) → relation filter → score+rank → assemble a context_prompt string. You inject context_prompt into your LLM's system prompt. Done. Quickstart Path A — Docker + Ollama (fully free, recommended) git clone https://github.com/zaydmulani09/mnemo cd mnemo docker compose up -d # Pull the llama3 model the first time (~4 GB) docker exec mnemo-ollama ollama pull llama3 # Verify everything is healthy curl http://localhost:8080/health Path B — Binary (Ollama or OpenAI running separately) cargo install --path crates/mnemo-api # With Ollama export MNEMO_LLM_BASE_URL=http://localhost:11434/v1 mnemo-api # With OpenAI export MNEMO_LLM_BASE_URL=https://api.openai.com/v1 export MNEMO_LLM_API_KEY=sk-... export MNEMO_LLM_MODEL=gpt-4o-mini export MNEMO_LLM_PROVIDER=openai mnemo-api Path C — Python SDK pip install mnemo-sdk from mnemo import MnemoClient client = MnemoClient () # server at http://localhost:8080 # Store a memory client . ingest ( "I'm building a Rust vector database called vecdb" ) # Get context for injection into your next LLM prompt print ( client . get_context ( "what am I working on?" )) API Reference All endpoints accept and return application/json . Base URL: http://localhost:8080 . Method Path Description Request body Response GET /health Server + DB + LLM status — HealthResponse POST /ingest Store text, extract entities IngestRequest IngestResponse POST /retrieve Retrieve ranked memory context RetrievalQuery RetrievalResult GET /entities List entities (paginated) ?limit&offset Entity[] GET /entities/:id Get entity by UUID — Entity DELETE /entities/:id Delete entity (cascades) — {"deleted":true} GET /entities/:id/neighbors Knowledge graph neighbors ?depth (max 5) GraphNode[] GET /chunks List memory chunks (paginated) ?limit&offset&session_id MemoryChunk[] GET /chunks/:id Get chunk by UUID — MemoryChunk DELETE /chunks/:id Delete chunk — {"deleted":true} POST /search Full-text search entities + chunks {"query","limit"} {"entities","chunks"} DELETE /wipe Delete all memory (irreversible) header: X-Confirm-Wipe: true {"wiped":true} GET /stats Entity/chunk/graph counts + uptime — StatsResponse Key request/response types: // IngestRequest { "content" : " string " , // required — text to store "source" : " string " , // required — e.g. "chat", "email", "cli" "session_id" : " string|null " , // optional — group related chunks "metadata" : {} // optional — arbitrary JSON } // RetrievalQuery { "text" : " string " , // required — query text "session_id" : " string|null " , // optional — filter by session "max_chunks" : 10 , // default 10 "max_entities" : 20 , // default 20 "min_confidence" : 0.5 , // default 0.5 "include_graph" : true , // default true — expand via knowledge graph "graph_depth" : 2 // default 2 — BFS depth for graph expansion } Full endpoint documentation with curl examples: docs/api.md Configuration Environment variables Variable Default Description MNEMO_DB_PATH mnemo.db SQLite database file path MNEMO_PORT 8080 API server port MNEMO_LLM_BASE_URL http://localhost:11434/v1 OpenAI-compatible LLM base URL MNEMO_LLM_MODEL llama3 Model name for entity extraction MNEMO_LLM_API_KEY ollama API key (any value works for Ollama) MNEMO_LLM_PROVIDER ollama Provider type: ollama , openai , anthropic , custom TOML config file Pass --config path/to/config.toml to mnemo-api . See mnemo.example.toml : db_path = " mnemo.db " port = 8080 [ llm ] provider = " ollama " base_url = " http://localhost:11434/v1 " model = " llama3 " api_key = " ollama " timeout_secs = 30 max_retries = 3 max_tokens = 2048 temperature = 0.1 Environment variables take precedence over TOML values. The active config source is reported in GET /health → config_source . CLI Install: cargo install --path crates/mnemo-cli Usage: # Store a memory mnemo ingest " I use Neovim and prefer dark mode " # Retrieve relevant context mnemo search " what editor do I use? " # List all extracted entities mnemo entities # Show entity detail + graph neighbors mnemo entity < uuid > --neighbors # List memory chunks mnemo chunks # Server health mnemo health # Memory statistics mnemo stats # Delete everything (prompts for confirmation) mnemo wipe # Skip confirmation prompt mnemo wipe --yes # Point at a non-default server mnemo --server http://192.168.1.10:8080 stats Python SDK Install: pip install mnemo-sdk See sdk/python/README.md for the full API reference. Async example: import asyncio from mnemo import AsyncMnemoClient async def main (): async with AsyncMnemoClient () as client : await client . ingest ( "Alice is a principal engineer at Stripe working on payment infrastructure." , session_id = "session-001" , ) context = await client . get_context ( "what does Alice work on?" , session_id = "session-001" , ) print ( context ) asyncio . run ( main ()) A working standalone example: examples/basic_usage.py Architecture Four Rust crates wired together: Crate Type Role mnemo-core lib Entity extraction, graph ops, retrieval engine, DB layer mnemo-api bin Axum REST API — thin handler layer over mnemo-core mnemo-cli bin CLI tool using blocking reqwest against the API mnemo-bench bin Performance benchmarks (12 suites) Full architecture documentation: docs/architecture.md Performance Benchmarked on Apple M2, SQLite WAL mode, in-memory petgraph. Debug build numbers — release build ( --release ) is 3–5× faster. Operation Avg latency Throughput Entity insert (SQLite) ~0.12 ms ~8,300 ops/s Entity lookup by ID ~0.08 ms ~12,500 ops/s Chunk insert ~0.14 ms ~7,100 ops/s Full-text chunk search ~0.28 ms ~3,500 ops/s Graph neighbor (depth=1) ~0.21 ms ~4,700 ops/s Graph neighbor (depth=2) ~0.89 ms ~1,100 ops/s Full retrieval pipeline ~4.2 ms ~238 ops/s Run cargo run -p mnemo-bench to benchmark on your hardware. Testing Rust cargo test --workspace # run all 122 tests make coverage # HTML coverage report (requires cargo-llvm-cov) make coverage-summary # summary to stdout Python SDK cd sdk/python && pytest tests/ -v Benchmarks cargo run -p mnemo-bench # all 12 benchmarks cargo run -p mnemo-bench -- --filter graph # graph benchmarks only cargo run -p mnemo-bench -- --json out.json # save results to JSON Current test counts: 122 Rust tests · 21 Python tests · 12 benchmarks Contributing PRs welcome. Please run make fmt && make lint before submitting. Open an issue first for large changes. See CONTRIBUTING.md for full setup instructions, code style guide, and how to add a new LLM provider. License MIT — see LICENSE