agentsclimarketplace
Sign in

Kage memory

Skill kage-core/Kage/skills/kage-memory

Trustworthy, code-grounded repo memory for coding agents. Before any task, recall grounded repo knowledge (decisions, bugs, conventions, code paths); after learning something durable, capture it; never act on stale or hallucinated memory. Use at the start of every task, when debugging, and after solving something reusable.From its SKILL.md

Install
npx -y skills add kage-core/Kage --skill kage-memory

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

2.3 KB, as published. Nobody here has run it

Kage — trustworthy repo memory

Kage gives this repo a memory that can be trusted: it rejects hallucinated citations on write, withholds stale memory on recall, and grounds knowledge to the code graph. Memory lives as plain files in .agent_memory/, reviewed in the same PR as the code.

One-time setup (if Kage isn't installed yet)

npm install -g @kage-core/kage-graph-mcp
kage init --project .
kage setup <your-agent> --project . --write   # claude-code, codex, cursor, windsurf, …

Try it instantly, no setup: npx -y @kage-core/kage-graph-mcp demo

How to use Kage during work

  1. Recall before acting. At the start of a task, debugging, or any repo-specific decision, call kage_context with the repo path and the task as the query. Use the returned memory only when it's relevant and source-backed. Prefer repo memory over your assumptions.
  2. Trust the gate. Kage withholds stale/ungrounded memory automatically — if something was withheld, don't go hunting for it; it's withheld because its evidence is gone.
  3. Capture durable learnings. When you fix a bug, make a decision, discover a convention or gotcha, or learn how to run/test/build the repo, call kage_learn with a concise learning and the file paths it concerns. Citations to files that don't exist are rejected — cite real paths.
  4. Keep memory honest on handoff. If you changed code that existing memory describes, update or supersede that memory (kage_verify_citations, kage_compact) before finishing.

What makes it trustworthy

  • Validated on write — memory citing nonexistent files is rejected.
  • Withheld on recall — memory whose cited code was deleted/expired is hidden.
  • Grounded to code — recall can return the bounded blast radius of a change.
  • Governed like code — packets are reviewable JSON in the repo.

Prove it: kage benchmark --trust --project .

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.