Graft

At a glance

Graft stores useful coding facts and past fixes locally so an agent can retrieve them when a later question uses different wording. A local embedding model turns note titles into searchable vectors, while text search covers titles and bodies. Notes live in SQLite databases separated by named profiles.

Good fit

A reasonable trial for recurring project questions that otherwise require rediscovering the same fix. Retrieval-shaped titles matter: the semantic index embeds the title, so “session cookies fail behind the proxy” is a more useful hypothetical title than “important note.” The body can preserve the explanation and commands that make the memory actionable.

Consider something else when

You need a stable production API or a guarantee that retrieved notes are correct. Graft is an alpha project with changing CLI and JSON interfaces. A strong semantic match means the note resembles the question; it does not establish that the advice applies to the present code.

Start here

After installing the documented CLI, choose a disposable profile and save one harmless fact with graft insert --title ... --body ... --keyword .... Query a paraphrase with graft query, then check the returned note against its source. Compare a relevant question with an unrelated one before making the profile part of an agent’s routine.

Cost and requirements

The Apache-2.0 tool runs a local daemon and uses a roughly 600 MB BGE-M3 model through llama.cpp. CPU execution is the default; CUDA and ROCm support are optional. Local memory needs no paid API, though the coding agent can still incur model charges. Separate profiles have separate databases and processes; choose them deliberately for different work.

Limits to know

Graft records source provenance and fingerprints, but its relevance labels still need judgment. Its published 0.1.1 benchmark used a small synthetic project and reported confident strong matches on 40% of negative retrieval questions. That older result is not a current-version accuracy estimate. Optional remote profile sync shares memory, and notes passed to an agent may also reach that agent’s model provider.

Related work

• Mnemolith: compare file-linked memories and checks for changed source files (alternative)