agent-knowledge

agent-knowledge

Give your coding agent a knowledge base that gets better over time.

agent-knowledge turns project documents, decisions, notes, and conversations into a portable Markdown wiki that your agent maintains for you. Ask a question and get a cited answer. Add a source and the agent integrates it with what the project already knows. Run a health check and it finds stale claims, contradictions, and orphaned pages before the wiki quietly rots.

Everything stays in your repository as plain Markdown + Git: readable without special tooling, diffable in code review, and portable across agents.

Install

Via skills.sh for Claude Code, Cursor, Codex, and 20+ other agents:

npx skills@latest add stjbrown/agent-knowledge

Or install it as a Claude Code plugin:

/plugin marketplace add stjbrown/agent-knowledge
/plugin install agent-knowledge

See it work

Start a knowledge base in any project:

/kb-init

Then use ordinary prompts:

Ingest this architecture decision: we chose Postgres because...

What do we know about authentication, and which sources support it?

What conflicts with our current deployment strategy?

The agent extracts durable knowledge, connects it to existing concepts, answers with citations, and files valuable new conclusions back into the bundle. Two explicit commands handle maintenance:

/kb-lint       # find broken links, stale claims, contradictions, and gaps
/kb-visualize  # explore the bundle as an interactive graph

The knowledge/ directory is a complete working example. It documents this project using the same format and skills the project provides.

Why this exists

Most agent "memory" is either retrieval over raw documents or a pile of notes that nobody maintains. The first repeatedly re-derives answers; the second gradually becomes untrustworthy. Neither makes knowledge stewardship an explicit job.

The hard part of a useful knowledge base is the bookkeeping: integrating new information, updating cross-references, preserving provenance, flagging contradictions, and keeping summaries current. That is exactly the work an agent can perform consistently.

agent-knowledge makes the agent a disciplined wiki maintainer:

• Ingest a source once → the agent extracts the signal and integrates it across the bundle.
• Query the bundle → it navigates by links, answers with citations, and files good answers back.
• Lint it → it catches drift (contradictions, stale claims, orphans) before the base rots.

Two design choices keep the result portable and trustworthy:

• A real, open format. Bundles follow Google's Open Knowledge Format (OKF) rather than a tool-specific database or hidden memory store.
• An explicit trust model. Meaning is append-only: the agent supersedes claims with provenance instead of silently rewriting history.

The workflow is based on Andrej Karpathy's LLM Wiki pattern, made conformant to OKF and packaged as skills you can drop into any project.

The skills

The family splits on who invokes them. Model-invoked skills the agent can reach for on its own when the task fits; user-invoked skills you trigger deliberately by name.

Model-invoked

• kb — the hub. Explains the format, holds the shared spec / glossary / trust model / templates, and routes to the right skill. Other kb-* skills read its reference as their single source of truth.
• kb-ingest — read a raw source once, extract its signal, and integrate it across the bundle under the trust model. The heart of the system.
• kb-query — answer a question from the bundle (or surface relevant context for another task) by progressive disclosure, cite the concepts used, and file valuable answers back so the base compounds.

User-invoked

• kb-init — scaffold a new bundle (default knowledge/, custom path, multi-bundle aware) and write its per-project schema layer (concept types + conventions) so the generic skills fit your domain.
• kb-lint — health-check the bundle: a deterministic OKF conformance pass plus a drift audit (contradictions, stale claims, orphans, coverage gaps), with an optional safe fix mode.
• kb-visualize — render the bundle as an interactive graph — native UI where the host supports it, otherwise a self-contained HTML file.

This repo documents itself in OKF