Yunaki Memory

A git-backed, per-repo memory layer exposed as an MCP server — the substrate for a self-evolving Claude skill. Drop the skill into any git repo and a dedicated, version-controlled memory store materializes on first write. Every memory is a commit, so the skill's growth is a readable, revertible history.

Install & use

The server is stdio-based: each person runs their own copy locally, against their own repos. Pick whichever install fits — none requires editing an absolute path.

npx from GitHub (no npm publish needed):

claude mcp add yunaki-memory -- npx -y github:Nanda-Kiran/yunaki-memory-mcp

npx from npm (after npm publish):

claude mcp add yunaki-memory -- npx -y yunaki-memory-mcp

Clone & build (zero publishing):

git clone https://github.com/Nanda-Kiran/yunaki-memory-mcp.git
cd yunaki-memory-mcp && npm install && npm run build
claude mcp add yunaki-memory -- node "$(pwd)/dist/index.js"

Equivalent .mcp.json (commit this in any repo to share the config with collaborators):

{
  "mcpServers": {
    "yunaki-memory": {
      "command": "npx",
      "args": ["-y", "github:Nanda-Kiran/yunaki-memory-mcp"]
    }
  }
}

Memory is stored per-user under ~/.yunaki/memory/<repo-id>/ (override with YUNAKI_MEMORY_ROOT). Installing the server does not share memories between people — see Roadmap for memory_sync.

Quick start

When you first adopt a repo, run memory_ingest once to seed memory with the repo's overview, file structure, and docs. After that the skill accumulates memories as it works.

How it works

• Identity, not path. Memory is keyed to a repo's root-commit SHA (git rev-list --max-parents=0 HEAD), which survives clone / rename / remote-move. Commit-less repos fall back to a UUID cached in .git/config. Path and remote URL are kept only as labels.
• Repo discovery. The server resolves the active repo from its working directory (the Claude Code CLI launches it inside your project), overridable per call via repoPath or pinned with the YUNAKI_REPO env var.
• Central sidecar. Stores live at ~/.yunaki/memory/<repo-id>/ — never pollutes the user's repo.
• Ad hoc creation. ensureRepoMemory(cwd) runs at the top of every tool call: resolves identity, and on first touch git inits the store, seeds repo.json + MEMORY.md, and commits.
• One file per memory. Markdown + YAML frontmatter under <type>/<slug>.md — clean diffs, trivial merges. MEMORY.md is the rebuilt index (the cheap working set).

Tools

Tool	Purpose
memory_ingest	Scan the repo and seed reference memories: overview (stack, scripts, file types), file tree, and docs. Idempotent.
memory_write	Persist a memory (fact/heuristic/failure/success/preference/reference) — commits it
memory_search	Retrieve memories for the current repo, ranked by keyword × confidence
memory_repo_info	Show resolved identity + memory location for the current repo

All accept an optional repoPath; otherwise they resolve against the server's cwd.

Develop

npm install
npm run build
npm run smoke           # core write/search loop
node test/ingest-smoke.mjs   # ingest against this repo

Roadmap (the "self-evolving" half)

1. memory_reinforce(id, outcome) — bump/decay confidence + usage_count instead of duplicating.
2. Dedup-before-write — semantic/slug match so reinforcement replaces near-duplicates.
3. memory_consolidate() — merge dupes, prune stale, and promote high-confidence memories into the skill's SKILL.md. That promotion (a git diff of the skill rewriting itself) is the payoff.
4. Embedding rerank (Voyage) cached in the gitignored .index/.
5. memory_sync() — push/pull each repo-memory git store to a remote for shared team memory.

License

MIT