Skip to content

CHANGELOG.md

Latest commit

 

History

History
83 lines (55 loc) · 4.73 KB

CHANGELOG.md

File metadata and controls

83 lines (55 loc) · 4.73 KB

Changelog

All notable changes to graft are documented in this file. The format follows Keep a Changelog and the project follows Semantic Versioning.

[Unreleased]

Changed

  • graft setup claudecode, graft setup codex, and graft setup opencode now install only the shared skills package. Hook installation and agent settings/instruction writes are disabled in code for the next version.

0.1.0 — Initial release

First public release of graft — local-first agentic memory for AI coding agents. A single binary + single SQLite file that gives any agent persistent memory across sessions, context resets, and machines, with no cloud and no API key.

Core engine

  • C11 daemon (graftd) with AF_UNIX socket transport and a thin graft CLI client; MessagePack wire protocol for low-overhead local IPC.
  • Storage on SQLite with sqlite-vec (dense vector index) and FTS5 (BM25 lexical index) in a single DB file.
  • Embeddings via embedded llama.cpp running BGE-M3 on CPU out of the box; opt-in GPU acceleration with GRAFT_GPU=cuda (NVIDIA) or GRAFT_GPU=hip (AMD ROCm 6 / 7).
  • WAL-safe sync layer: single-writer push half, no double-open, no WAL contention under concurrent reads.

Retrieval

  • graft query — verified semantic cache returning STRONG / WEAK / MISS in milliseconds. Multi-signal verifier refuses to claim a hit when dense and lexical signals disagree, so agents never quote confidently-wrong answers.
  • graft retrieve — hybrid search fusing dense (BGE-M3 cosine) and lexical (BM25 over title + body) via Reciprocal Rank Fusion.
  • graft explore — beam-search graph walk over keyword and semantic edges with MMR diversity and gamma^step decay.

Knowledge model

  • Memory nodes: title (retrieval anchor) + body (full context) + keywords.
  • Graph edges: keyword and semantic links between nodes, walked by explore.
  • Supersession: replace outdated nodes atomically while keeping the old version visible as SUPERSEDED — history stays, mistakes don't propagate.
  • Confidence levels surfaced to clients (STRONG / WEAK / MISS) so agents can gate behavior on retrieval quality.

Multi-tenant profiles

  • Isolated DBs and sockets per profile (work, personal, project-scoped); switch with GRAFT_PROFILE=<name>.
  • Import / export / merge profiles as plain SQLite files — portable, diffable, scriptable.

Optional REST API + 3D viewer

  • Nine JSON endpoints (/v1/match, /v1/search, /v1/insert, …) gated by a flag in config.yaml.
  • Browser-based 3D graph viewer with click-to-edit and atomic supersession.

Agent integrations

Each adapter ships skills (when to search, when to save) and, where the harness supports them, optional hooks (deterministic execution on UserPromptSubmit / PostToolUse / Stop so the model can't forget):

  • Claude Code — skills + optional hooks.
  • Codex — skills + optional AGENTS.md and hooks.
  • Claude Desktop — MCP server (stdio).
  • ChatGPT — MCP server (stdio or HTTP) with optional OAuth gateway.
  • Gemini CLI — GEMINI.md memory file.
  • Open Code — skills + optional AGENTS.md.

Microservices pattern

  • Reference architecture for L1 Redis + L2 graft semantic cache + L3 graft + LLM with writeback, documented in docs/microservices/.
  • Designed to absorb most "GPT in a microservice" traffic before it hits the LLM, with the system getting cheaper and faster over time as L3 answers write back into L2.

Packaging & distribution

  • Homebrew formula (Formula/graft.rb) with prebuilt Linux bottle.
  • Scoop bucket (bucket/graft.json) for Windows users.
  • Cross-platform installer scripts: scripts/install.sh (Linux / macOS / MSYS2) and scripts/install.ps1 (Windows, auto-installs MSYS2).
  • Build under 3 minutes on a laptop.

CI / release pipeline

  • Single unified release workflow triggered by pushes to the release branch.
  • Strict semver gate (only x.y.z, no pre-release suffixes).
  • ctest → parallel build of Linux tarball, Homebrew bottle, Scoop zip → single signed publish job with cosign signatures, SBOM, and build provenance attestations.
  • Smart caching of llama.cpp build (keyed on submodule SHA) and ccache across runs.

Documentation

  • Per-feature docs tree under docs/ (install, retrieval, insert, profiles, embeddings, integrations, microservices, HTTP API, viewer, maintenance, storage, architecture, release).
  • Glossary in docs/concepts.md, use cases in docs/use-cases.md.
  • Contributor guide in CONTRIBUTING.md with commit-msg policy (Conventional Commits, English ASCII subject ≤ 70 chars).