🧠 100% local · MCP-native · MIT-licensed

Memory your AI agent actually keeps

Local, private long-term memory for AI coding agents — hybrid graph + vector search, two-stage verification, and an autonomous Night Gardener. Runs entirely on your machine. Zero cloud by default.

🚀  Quick Start ★  Star on GitHub

uv tool install hydramem
hydramem init ~/my-memory && cd ~/my-memory
hydramem ingest ./kms && hydramem search "what did we decide about auth?"

0token savings vs naive RAG
auditable, not a vibe

0MCP tools
Claude Desktop · Cursor · OpenCode

0lines of code
readable in an afternoon

0bytes sent to the cloud
by default

$ hydramem stats --last-7d
╭─────────────────────────────────────────────────╮
│  HydraMem Stats – last 7 days                   │
├──────────────────────────┬──────────────────────┤
│ Tool calls               │                 142  │
│ Tokens (naive RAG)       │               1.4M   │
│ Tokens injected          │               312K   │
│ Tokens saved             │    1.09M (77.8%)     │
│ Avg VoG score            │              0.887   │
│ Rejected by SR-MKG       │                 89   │
│ Hallucinations blocked   │                  5   │
╰──────────────────────────┴──────────────────────╯

Why HydraMem

•  100% local, zero exfiltration

  ---

  Your codebase context, graph, vectors and telemetry never leave your machine. No cloud memory service in the loop. Secrets come from env vars only.

•  Hybrid graph + vector search

  ---

  LanceDB / Grafeo vectors + graph traversal + BM25, fused with Reciprocal Rank Fusion in a single query.

•  Two-stage verification

  ---

  Every relation passes a topological SR-MKG scorer and an optional VoG groundedness check — so the graph never fills with hallucinated edges.

•  Autonomous Night Gardener

  ---

  Offline relation inference & pruning that refines the graph overnight — and emits zero relations when there is no real evidence.

•  MCP-native (18 tools)

  ---

  A FastMCP server that plugs into Claude Desktop, Cursor and OpenCode over stdio or HTTP. Single-tenant by design.

•  Honest & auditable

  ---

  hydramem stats --raw audits every token saved. No fabricated metrics — if a component doesn't measurably work, the docs say so.

Quick start in 60 seconds

uv (recommended)pip / pipxFrom source

uv tool install hydramem
hydramem init ~/my-memory
cd ~/my-memory

pip install hydramem
# …or: pipx install hydramem
hydramem init ~/my-memory
cd ~/my-memory

git clone https://github.com/xusliebana/hydramem && cd hydramem
cp config.yml.example config.yml
uv sync
uv run hydramem --help

Then ingest your notes and ask a question — everything runs locally:

hydramem ingest ./kms --project myproject
hydramem search "what does my documentation say about the Night Gardener?"
hydramem serve --transport stdio        # expose to your MCP client

Full quick start → All MCP tools → Integrations →

How it works

AI client (Claude Desktop / Cursor / OpenCode)
        │  MCP (stdio · http)
        ▼
HydraMem MCP server (FastMCP · 18 tools · local telemetry)
        │
   ┌────┴───────────────┬────────────────────┐
   ▼                    ▼                     ▼
Ingest             Hybrid search         Night Gardener
chunk → embed      vector + graph        infer → verify → prune
(Nomic v1.5,       + BM25 (RRF)          (offline, evidence-only)
 512-d, local)     → SR-MKG → VoG
        │                    │                    │
        └──────────┬─────────┴──────────┬─────────┘
                   ▼                     ▼
        Graph store (Grafeo)   Vector store (LanceDB / Grafeo HNSW)

The default stack is Gemma 4 (gemma4:e4b) for local reasoning/verification and Nomic Embed Text v1.5 (Matryoshka, truncated to 512-d) for embeddings — both run locally. Swap any of them in config.yml.

How does it compare?

Project	Local-first	Graph + vector	Verifies relations	Offline learning	MCP-native	License
HydraMem	✅	✅	✅ (SR-MKG + VoG)	✅ (Night Gardener)	✅	MIT
Mem0	partial	optional	❌	partial	community	Apache-2.0
Letta / MemGPT	✅	❌	❌	❌	❌	Apache-2.0
Zep / Graphiti	partial	✅	partial	✅	community	Apache-2.0
MS GraphRAG	✅	✅	indirect	❌	❌	MIT

Honest positioning

HydraMem is not the first memory system, graph-RAG, or hallucination filter. It's an opinionated local-first integration of those patterns, small enough to read in an afternoon (~5k LOC) and with metrics you can audit.

Give your agent a memory that stays yours.

🚀  Get started �  Integrations �📦  MCP tools reference ★  Star on GitHub