graphCTX

Local-first memory for coding agents.

Lifecycle-hook push for commit-valid, scope-aware context in Claude Code, plus refreshed static grounding and MCP/recall fallback for Cursor, OpenCode, and generic clients.

---

Using this from an AI agent? Start with AGENT_USAGE.md. It covers install, MCP server setup for every major client (Claude Code, Cursor, OpenCode, Codex, Claude Desktop, Zed, Continue, etc.), how to add graphCTX as a skill, and the 8 MCP tools with examples.

Why graphCTX

Coding agents forget. Compaction strips context, branches mutate truth, and "recall" only works when the model remembers to ask. graphCTX flips the model: it pushes commit-valid memory at the right lifecycle moments, and falls back to pull-based MCP recall where push is not available.

Everything runs on your machine. The CLI, embedded SQLite store, git anchoring, eval harness, benchmarks, and MCP stdio server work fully offline. There is no required web service and no required network call for normal operation.

Features

• Commit-valid fact store over SQLite with migrations, WAL, FTS, optional sqlite-vec, provenance edges, and append-only audit records.
• Hybrid retrieval with lexical, semantic, RRF, MMR diversity, scope and entity signals, and commit-valid filtering.
• Injection planner with lifecycle gate, token budgeting, anti-repetition ledger, open-loop resurfacing, send-edge security checks, and provenance tags.
• Multi-client adapters for Claude Code (lifecycle push), Cursor, OpenCode, generic clients, and proxy channel fallback for refreshed grounding plus MCP recall.
• MCP stdio server exposing exactly 8 tools: remember, recall, inject_context, checkpoint_session, promote, forget, why, resolve_conflict.
• Deterministic offline eval suite with benchmark gates covering memory, retrieval, relevance, security, temporal validity, conflicts, LLM extraction, promotion, adapters, MCP, storage, telemetry, provenance, resilience, docs, demo, and code quality.

Install

One-liner. Downloads a single self-contained binary for your platform (Linux or macOS, x64 or arm64). No Node, no Bun, no account, no API key. The binary bundles SQLite and the vector index, and no data leaves your machine.

curl -fsSL https://graph.coder.company/install | sh

Environment overrides:

• GRAPHCTX_INSTALL_DIR — install location (default: $HOME/.local/bin).
• GRAPHCTX_RELEASE — pin to a specific release tag.

Verify:

graphctx --version

From source (development)

Requires Node.js 20+.

git clone https://github.com/coder-company/graphCTX.git
cd graphCTX
npm install
npx tsx src/cli.ts --help

Compile the package when you need a build:

npm run build

For local CLI development either use npx tsx src/cli.ts ... directly or drop a graphctx shim on your PATH:

printf '#!/usr/bin/env bash\nexec npx tsx %s/src/cli.ts "$@"\n' "$PWD" > ~/.local/bin/graphctx
chmod +x ~/.local/bin/graphctx

Core Workflow

Initialize a repo, install the adapter, and inspect readiness:

graphctx init -C .
graphctx install auto -C .
graphctx doctor -C .

Store and retrieve explicit memory:

graphctx remember "deploy with ./scripts/ship.sh" -C .
graphctx recall   "how do I deploy this project"  -C .
graphctx why      <fact_id_or_last8>              -C .

Carry unfinished work across compaction:

graphctx loop    "finish the retry backoff"  -C .
graphctx hook    PostCompact -C . < payload.json
graphctx resolve <fact_id_or_last8> -C .

Run extraction and the offline demo:

graphctx extract -C .
graphctx demo

CLI Reference

Core installed commands are discoverable with graphctx --help:

graphctx init
graphctx install   <claude|cursor|opencode|generic|auto>
graphctx uninstall <claude|cursor|opencode|generic>
graphctx hook      <event>
graphctx recall    "<query>"
graphctx remember  "<text>"
graphctx loop      "<text>"
graphctx resolve   <fact_id_or_last8>
graphctx extract
graphctx serve     --mcp
graphctx why       <fact_id_or_last8>
graphctx doctor
graphctx demo
graphctx skill   <claude|cursor|opencode|codex|generic|all>
graphctx tui

Development-checkout gates:

npx tsx src/cli.ts eval all
npx tsx src/cli.ts eval quality
npx tsx src/cli.ts eval cli-docs-demo
npx tsx src/cli.ts eval mcp
npx tsx src/cli.ts eval run --arms A,B,C,N,S
npx tsx src/cli.ts bench

The development-only command names are still exposed by the CLI for repo checkouts: graphctx eval <suite|all>, graphctx bench, and graphctx compare.

Repository Layout

.
├── src/          # TypeScript source (CLI, store, retrieval, adapters, MCP)
├── test/         # Vitest unit/integration suites
├── fixtures/     # Repo fixtures used by extractors and eval arms
├── scripts/      # Build, packaging, and pack-smoke helpers
├── docs/         # Spec, gameplan, PRD, status, notes
├── DEMO.md       # ~3 minute live demo walkthrough
└── README.md

Testing and Quality Gates

Run the same green bar used by the development mission:

npx tsc --noEmit
npx biome check src test
npx biome check .
npx vitest run
npm  run   pack:smoke
npx tsx src/cli.ts eval all
npx tsx src/cli.ts bench

Expected current state:

• npx vitest run: 286 tests pass.
• npm run pack:smoke: packed tarball installs, runs the demo, and serves MCP from a clean temp app.
• npx tsx src/cli.ts eval all: 19 gate suites pass.
• npx tsx src/cli.ts bench: hook hot-path p95 below 150ms.

Design Notes

graphCTX is deliberately local-first. Missing LLM keys, missing git metadata, bad config, or corrupt stores degrade to no memory rather than a broken agent. Live LLM extraction is opt-in: deterministic extraction and all default gates run offline. Provider-backed extraction uses bounded, cancellable requests and fails soft to deterministic-only behavior.

The product thesis is measured by the offline ablation:

graphctx eval run --arms A,B,C,N,S

Arm C (lifecycle push) must beat arm B (pull-only recall), while the N/S controls prove graphCTX can deliver a memory-only fact and suppress stale facts.

For AI agents

If you are an AI agent (Claude Code, Cursor, OpenCode, Codex, Claude Desktop, Zed, Continue, a custom MCP client, etc.) read AGENT_USAGE.md. It is the single source of truth for:

• how to install graphCTX
• how to add the MCP server (per-client snippets)
• how to wire lifecycle hooks
• how to register graphCTX as a skill
• the 8 MCP tools and when to call each
• troubleshooting

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for development setup, the local green-bar checklist, and the conventions used by this repository. To report a security issue privately, see SECURITY.md.

License

MIT (c) coder-company