Graft — a local-first semantic memory system for AI agents that stores and retrieves learned context across sessions without cloud services.

Graft is a single-binary, SQLite-based memory layer for AI agents and microservices. It runs locally with no SaaS, no API key, and no Python runtime required; models (BGE-M3 via llama.cpp) run on CPU by default, with optional CUDA or ROCm GPU acceleration.

The system uses hybrid search (dense semantic embedding via BGE-M3 cosine similarity plus lexical BM25) fused by Reciprocal Rank Fusion, returning STRONG / WEAK / MISS verdicts in milliseconds. Results are gated by a verify step to prevent confident false matches; graph walks follow keyword and semantic edges using beam search and MMR diversity.

In a three-layer microservice stack, graft serves as L2 semantic cache (~30–80 ms latency) between L1 Redis exact-match caching (~1 ms) and L3 LLM + agentic retrieval (~500 ms+). Every L3 answer is written back via POST /v1/insert, so repeat queries hit the faster cached layer.

Integrations include Claude Code (skills + hooks), Codex (AGENTS.md + hooks), Claude Desktop and ChatGPT (MCP), and Gemini CLI. The C11 codebase (~10 K LOC project code) builds in under 3 minutes; multi-tenant profiles (work, personal, project-scoped) are isolated as separate SQLite databases and daemons.