Case studyAuthor and maintainer2026-08 – present

accreta

A knowledge base an AI agent writes and maintains, where every claim cites its source down to the revision and a drift command says which pages a source change has made suspect.

As of September 2026.

6/6planned phases shippedYou can check this: accreta, accreta on npm.
≈1 mofirst commit to a published packageYou can check this: accreta, accreta on npm.
4releasesYou can check this: accreta, accreta on npm.
npminstallable by anyone; atlas installs it from the registryYou can check this: accreta on npm, accreta-atlas.

How it works

Diagram generated from the knowledge base
How accreta works Sources pinned to a revision are compiled by an agent, under its versioned rules, into wiki pages whose claims cite line range and revision. The pages are indexed in SQLite FTS5 and served to people through a CLI and to agents through an MCP server. When a source moves, drift names the pages now suspect. a source moves → drift names the pages now suspect Sourcesrepos, docs, RFCs,pinned to a revision Agent compilesunder the constitution,its versioned rules Wiki pagesevery claim cites its linerange and revision IndexSQLite FTS5 CLIa person at a terminal MCP serverany agent

Context

Giving an agent knowledge of a large system usually means RAG: embed everything, retrieve chunks per question, hope the model assembles them. The context is rebuilt from scratch every time and nothing learned in one session survives into the next.

Problem

A retrieval index is not an artefact that improves. Cross-references, synthesis and contradictions are re-derived on every query, at cost, and never written down. Worse, documentation of any kind goes stale silently: nothing tells you which pages a change to the underlying system has just invalidated.

Role

Author and maintainer. Design, implementation and release.

Decisions

01 · Knowledge engineering for agentsCompile knowledge into a wiki, do not retrieve it per queryKnowledge engineering for agents

Following the LLM wiki pattern, the agent compiles knowledge once into interlinked Markdown pages and then keeps them current. The knowledge base becomes a compounding artefact rather than a lookup index, so cross-references and contradictions are already written down when a question arrives.

Ask: Why compile a wiki instead of using RAG?
02 · Knowledge engineering for agentsProvenance down to the revisionKnowledge engineering for agents

Every non-trivial claim cites the source it came from, including the line range and the revision. A page without citations is a page under suspicion. This is what makes the next decision possible.

Ask: Why cite sources down to the revision?
03 · Knowledge engineering for agentsDrift detection as a first-class commandKnowledge engineering for agents

Each page records the source revision it was last verified against, so when a source moves, drift names the pages that are now suspect. Documentation does not do this on its own, and it is the difference between a knowledge base and a pile of stale prose.

Ask: How does it know a page has gone stale?
04 · MCP serversShip an MCP server beside the CLIMCP servers

The same core is reachable by a person at a terminal and by any agent speaking MCP, so search, page fetch, canonical lookup and impact analysis are available without a bespoke integration.

Ask: Why ship an MCP server as well as a CLI?

Details

Skills
Agent-assisted deliveryContext engineeringTypeScriptMCP serversKnowledge engineering for agents

Results

  • Published on npm; all six planned phases complete and the pipeline runs end to end, with a worked example in the repository.
  • A companion repository exercises it against the IETF RFC series, installed from the registry, with drift running in CI against fixtures designed to make it go red.
  • The API is explicitly not yet stable.
  • Published on npm and installable by anyone; the companion repository installs it from the registry rather than from source, so a broken package fails there.

    you can check this: accreta on npm, accreta-atlas

  • First commit to a published package with all six planned phases complete in about a month, across four releases.

    you can check this: accreta, accreta on npm

    How was it built so quickly?