Salem J. AbabnehHow I workMethod 06

Giving an organisation a memory its agents can build on

Most organisations keep knowledge where software can barely reach it: slide decks, inboxes, someone's head. If knowledge instead lives as indexed, version-controlled files, then the same memory serves the people doing the work and the agents doing it with them.

Team memory systems compiling into one corporate memory Each team compiles its own raw sources into a file-based memory with its own index. Those merge into a single corporate memory, which both people and agents read from and write back to. Raw sourcesRaw sourcesRaw sourcesTeam memoryfiles + indexTeam memoryfiles + indexTeam memoryfiles + index One corporate memory indexed, cross-linked, versioned in git People read it Agents build on it

The method

  1. Knowledge becomes files, with an index

    Plain text, cross-linked, in version control, with an index a reader traverses instead of searching blindly. Andrej Karpathy published this shape in April 2026 as the LLM wiki: raw sources kept immutable, a compiled knowledge layer maintained on top, and a schema that says how the two relate.

  2. Compile knowledge, do not just retrieve it

    The distinction that matters. Conventional retrieval re-reads raw documents on every question and rebuilds the same understanding each time. A compiled memory writes the conclusion down once, links it, and improves it as new sources arrive, so the knowledge gets better rather than merely more numerous.

  3. Every team keeps its own memory

    A team's memory is the team's, in its own language, at its own granularity. Forcing one schema across an organisation on day one is how knowledge programmes die.

  4. Then merge them into one corporate memory

    Team memories compose upward into a single organisational memory, indexed and cross-linked. That merged layer is the asset: the institution's knowledge in a form both a new joiner and an agent can traverse on their first day.

  5. Contradictions surface instead of hiding

    When a new source disagrees with what is written, it gets flagged rather than silently overwritten. In most organisations two teams simply hold different truths and nobody finds out until it costs something.

  6. Agents become participants, not tools bolted on

    Once the memory is machine-traversable, an agent does not need a bespoke integration to be useful. It reads what the organisation knows, does work against it, and writes back what it learned. That is what makes a company AI-augmented at its core rather than AI-flavoured at the edges.

Where this comes from

This is not a proposal. The assistant that helps run my own work is built exactly this way and has been for over a year: an indexed memory of decisions, corrections and project state in version control, read and written by both me and the agents working alongside me. Every rule it follows was written the day something was learned. The methods on this site were drafted from that memory.

The other methods

Changing how a large organization works? Let's talk.

[email protected] LinkedIn Download CV