my-memory-wiki

Git-synced, agent-maintained team memory for Claude Code.

Claude captures durable knowledge — decisions, fixes, library gotchas, hard rules — as you work. It lives as Markdown in your repo under .memory/, is shared through git, and is gated by PR review. The knowledge stays with the code; this package is the engine (a CLI + an MCP server).

Built on Karpathy's LLM Wiki pattern and the consolidation ideas from Akita's ai-memory, adapted for teams: git + PR review are the consolidation gate.

Quick start

npx my-memory-wiki init     # scaffold .memory/ + CLAUDE.md + .mcp.json into the repo

Commit the result. Every teammate now gets the same memory on git clone, and their Claude Code reads it automatically — no per-developer setup.

How it works

Two tiers.

• Tier 0 — constraints. Hard rules ("never edit generated files"), always loaded into Claude's context via @.memory/CONSTRAINTS.md imported in CLAUDE.md. Strong default — Claude honours them unless you override in-session.
• Tier 1 — knowledge. Decisions, conventions, library notes, and reports. Retrieved on demand by skimming index.md or via the memory.search MCP tool.

The loop. Claude detects something durable (you state a rule, explain a non-obvious choice, root-cause a bug, or finish an investigation), writes a page and tells you in one line, and it lands in your PR. Merge is the review gate; CI lint catches conflicts.

What gets remembered is defined by the rubric in .memory/SCHEMA.md (criteria + calibration examples + a project-triggers section you customize). Spontaneous capture is prompt-driven via the CLAUDE.md/AGENTS.md loop; for a deliberate pass, run the /scribe slash command (or my-memory-wiki scribe) — it reads the live SCHEMA.md and reviews the conversation against it.

Source vs. views. Page frontmatter (YAML) is the source of truth. index.md (human/LLM), index.json (machine), and CONSTRAINTS.md (always-loaded) are all generated — run my-memory-wiki index (or let the MCP remember tool do it).

CLI

| Command | What it does | |---------|--------------| | my-memory-wiki init | Scaffold .memory/, the CLAUDE.md/AGENTS.md sections, .mcp.json, and (when .github/ exists) a CI lint workflow. Idempotent; --no-ci to skip the workflow. | | my-memory-wiki index | Regenerate index.md, index.json, CONSTRAINTS.md from pages. | | my-memory-wiki lint [--ci] | Check for broken links, conflicting rules, staleness, drifted views. --ci exits nonzero on errors. | | my-memory-wiki serve | Run the MCP server over stdio (started automatically by Claude Code). | | my-memory-wiki update | Refresh package-managed files to the installed version, preserving your edits (writes SCHEMA.md.new on conflict). | | my-memory-wiki scribe | Print the rubric-driven prompt for a memory-capture pass (also scaffolded as the /scribe Claude Code slash command). |

MCP server

my-memory-wiki init writes a .mcp.json that registers the server. Claude Code spawns it automatically per session (local stdio, no deploy, no daemon). It exposes:

• memory.search — find pages by meaning (BM25)
• memory.remember — record a valid page + regenerate views + log
• memory.related — a page's links, backlinks, same-tag pages, supersedes chain
• memory.lint — the lint rules, mid-session

You don't need the server for small wikis — the always-on constraints and reading .memory/ directly are enough. It earns its place once index.md is too big to skim.

CI

Add my-memory-wiki lint --ci to your CI so a PR can't merge with broken links, conflicting constraints, or out-of-date views. See docs/install.md.

License

MIT