Let AI work with your personal context

Keep is your personal context library for thoughts, ideas, notes, links, saved content, and things you learn. The Keep skill teaches compatible AI tools when and how to search that library and save useful context back to it.

Install

If you use the Keep CLI:

keep skill install

Or install directly from Keep:

npx skills add https://keep.md/.well-known/agent-skills/keep

yarn dlx skills add https://keep.md/.well-known/agent-skills/keep

pnpm dlx skills add https://keep.md/.well-known/agent-skills/keep

bunx skills add https://keep.md/.well-known/agent-skills/keep

For non-interactive installs (CI or automation):

npx skills add https://keep.md/.well-known/agent-skills/keep --all

yarn dlx skills add https://keep.md/.well-known/agent-skills/keep --all

pnpm dlx skills add https://keep.md/.well-known/agent-skills/keep --all

bunx skills add https://keep.md/.well-known/agent-skills/keep --all

The install includes the skill instructions and its OpenAI interface metadata.

Check for updates at any time:

keep skill status
keep skill update

What AI can do

With the skill installed, compatible AI tools can:

• recognise recall cues such as “I saved a site a few months ago” and search without needing an exact title or URL
• proactively search project context when an earlier decision, handoff, preference, or piece of research could affect the work
• search notes and saved items together with semantic and keyword matching
• read attached sources, highlights, summaries, and bounded full content
• create or update notes for thoughts, ideas, decisions, preferences, learning, and handoffs
• capture a promising idea, such as a future blog post, without copying the whole conversation
• attach saved items as evidence instead of copying whole articles into notes
• process inbox items and work with automated sources and sync workflows

The skill does not search on every ordinary request or save every session. It saves useful durable context, not full chat transcripts, secrets, temporary logs, generic facts, or unsupported guesses.

The context loop

The skill gives each compatible AI client the same small workflow:

1. Recall context when you refer to something saved or discussed before, or when project-specific knowledge could materially affect the work.
2. Open only the Notes or Feed Items that matter.
3. Search again before saving, then improve an existing Note or create one.
4. Before finishing substantial work, consider zero to three durable outcomes worth carrying forward.

Zero saved outcomes is normal. Keep prefers one living Note per topic or project over a new session summary every time.

Notes are editable knowledge. Feed Items are captured source material. Adding commentary to an Item creates or updates a Note and links the Item as evidence rather than changing the captured content.

The skill uses keep search for Notes and Items together. It uses keep notes search or keep items search when it needs one resource type. keep context is only a compatibility alias.

Search returns a small map, not whole documents. When a Note result includes a useful line range, the AI can fetch that revision-pinned slice directly. For a broader match, it opens the overview, searches within that revision, then fetches an exact slice. The complete body is the last resort.

Each connected agent uses the same library. A named credential records which client changed a Note. Revision checks stop a stale client from silently overwriting a newer change. See Connected agents.

Requirements

Use the authenticated Keep CLI where a terminal is available, or connect the Keep MCP server to a compatible AI client. Both provide access to the same Keep library.

Use session hooks when Claude Code, Codex, or Pi should load a small project index at session start and save a handoff after a completed session. Cursor has a start-only preview, and OpenCode has an explicit handoff command.