Home

graft is a single-binary semantic cache that survives across sessions. Agents save what they learn; queries return verified hits in milliseconds; misses degrade to graph exploration. No SaaS, no API keys, one SQLite file.

LLM/agent ──> graft query "…"
                 │
                 ▼
            STRONG  ──> reuse answer (no LLM call, no token cost)
            WEAK    ──> reuse with low-confidence banner
            MISS    ──> graft retrieve / explore fallback
                 │
                 ▼
            graft insert  ──> next time it's a STRONG

What graft is

A C11 daemon + CLI that speaks MessagePack over an AF_UNIX socket, backed by SQLite + sqlite-vec + FTS5, embeds text with BGE-M3 (1024-dim) via llama.cpp, and exposes both a subprocess CLI and an optional REST + 3D viewer. Designed to be the smallest useful thing that makes hard-won agent knowledge survive a session.

Why not a vector DB / RAG / chatbot platform

	Vector DB	RAG framework	graft
Local-first, no SaaS	✗	partial	✓
One binary, no dependencies	✗	✗	✓
Graph topology (not just lists)	✗	✗	✓
Verified hits (STRONG/WEAK/MISS)	✗	✗	✓
3D viewer included	✗	✗	✓
Multi-tenant by profile	partial	✗	✓
MCP + Skills/Hooks integrations	✗	partial	✓

Where to next

• First time? Start at Quick Start.
• Want the big picture? Read Architecture.
• Looking up a command? See CLI Reference or REST API.
• Tuning? Every knob is in Configuration.
• Stuck? Try Troubleshooting.