The Agent Wiki Pattern

Most systems still give models knowledge the hard way. They retrieve raw chunks at query time and force the model to re-derive understanding on every question. Nothing accumulates.

A better pattern has quietly emerged in 2026.

An LLM reads a body of sources once, compiles them into a maintained set of markdown pages, and keeps those pages current as new material arrives. Agents then read the pages instead of starting from raw documents every time.

The architecture is consistent across independent implementations:

•⁠ ⁠Raw sources stay immutable

•⁠ ⁠The wiki is LLM-generated and model-owned

•⁠ ⁠A schema file tells the model how to organize and maintain the pages

Three operations run on top: ingest, query, and lint.

This works because language models do not mind the maintenance work that kills human wikis — updating cross-references, reconciling contradictions, and keeping summaries current.

The pattern has clear limits. It works best at moderate scale. Early summaries can quietly lose detail. Stale pages become confidently wrong. And a wiki is not memory. It compiles what a body of material contains. It does not remember what a specific user prefers, decided, or already rejected.

This Ledger Entry expands how the reader thinks about agent context systems by clarifying a rising architectural pattern: compile knowledge once at ingest time and maintain it, rather than re-deriving it on every query. It also draws a useful line between corpus knowledge and true user memory.