MEX keeps a team's architecture notes, decisions, requirements, and handoffs in the repository as structured Markdown, where both engineers and their coding agents can read and extend them. Git is the sharing layer: canonical files travel with ordinary commit, push, and pull, while each teammate keeps their own local indexes, drafts, identity selection, and Hub. No hosted service, Docker, proxy, account, or model key is required.
Two complementary views of a repository back the memory. The Wiki explains architecture, conventions, decisions, patterns, topics, and Specs in reviewable language. The Code Graph uses bundled Tree-sitter grammars to map symbols and relationships from the implementation into a local SQLite index for bounded retrieval, covering TypeScript/TSX, JavaScript/JSX, Python, and Rust, with partial C# support and framework resolvers for Express, Next.js App Router, FastAPI, Flask, and NestJS. A Wiki claim can be grounded to a graph node by ID, fingerprint, and body hash, so a later code change flags the claim as changed, moved, missing, ambiguous, or unverified. Drift is a review signal, not proof that the prose is wrong.
Agents act as readers and contributors. Setup installs short host-agent instructions that point at a policy file and a router file, through which Claude Code and Codex retrieve task-relevant context and run the Inbox and Relay skills. Cursor, Windsurf, GitHub Copilot, and OpenCode receive an instruction anchor without a skill command. An agent can draft an Inbox proposal (an addition or correction to Wiki knowledge) or a Relay (a handoff with progress, decisions, blockers, evidence, and next actions), but a person reviews and publishes every canonical write in the local Project Hub.
The Hub is a loopback web app per checkout, not a shared dashboard, with pages for search, the Context graph, code, Inbox, Relays, members, activity, and index health. A terminal dashboard and a plain CLI cover the same workflows, and a separate agent-memory mode applies the routing and maintenance model to long-running operational agent workspaces. An MCP server exists in the source tree only and is not a released interface.
Features
- Repository-native memory: architecture, decisions, specs, and handoffs stored as structured Markdown with metadata, relations, provenance, and code groundings, shared through Git
- Code Graph: Tree-sitter extraction of symbols and relationships into a local SQLite index, queried for task-scoped evidence, definitions, callers, and impact
- Grounding and drift detection: Wiki claims link to graph nodes by ID, fingerprint, and body hash, so code changes flag claims as changed, moved, missing, or ambiguous
- Inbox proposals: agents or people propose additions and corrections to project knowledge for explicit human approval
- Relays: structured handoffs carrying progress, decisions, blockers, evidence, and next actions from one engineer to the next
- Project Hub: a local, per-checkout web app for exploring the Context graph, reviewing Inbox and Relays, selecting identity, and running index maintenance
- Agent integrations: Claude Code and Codex get project instructions plus Inbox and Relay skills, and Cursor, Windsurf, GitHub Copilot, and OpenCode get an instruction anchor
- Language coverage: TypeScript/TSX, JavaScript/JSX, Python, and Rust, partial C#, with bounded resolvers for Express, Next.js App Router, FastAPI, Flask, and NestJS
- Incremental refresh: graph refresh reuses unchanged extractions and runs in a disposable process that releases compiler memory on exit
- Agent-memory mode: a separate template that applies the routing and maintenance model to homelab, infrastructure, and long-running agent workspaces
- Local-only trust model: no canonical records, indexes, drafts, or sessions are uploaded, and MEX never pushes or pulls on its own
