LLM Wiki
A personal knowledge base that an agent compiles and keeps current, instead of a pile of documents an agent re-reads from scratch on every question.
CREATED JUN 16, 2026 · UPDATED AUG 20, 2026
PROBLEM
Retrieval-augmented search never accumulates. Ask a question that needs five sources synthesised and the model finds and stitches the same fragments it stitched last week — nothing is built up, nothing is remembered, and the contradiction between source three and source nine goes unnoticed every single time.
HYPOTHESIS
If the agent writes the wiki instead of querying the corpus, knowledge compiles once and is then maintained. Cross-references are already there when you arrive; contradictions were flagged on ingest, not rediscovered on read.
ARCHITECTURE
Three layers. Raw sources are immutable and the agent only reads them. The wiki is a directory of interlinked markdown the agent owns outright — entity pages, concept pages, an index, an append-only log. The schema file is the configuration that makes the agent a disciplined maintainer rather than a chatbot, and it co-evolves with the vault. Three operations: ingest (one source can touch fifteen pages), query (and good answers get filed back as new pages, so exploration compounds too), and lint (contradictions, stale claims, orphans, missing pages).
STACK
RESULT
The pattern document is written and the vault is being built against it. A Next.js reader lives in the same repo.
LEARNINGS
- The expensive part of a knowledge base was never the reading — it was the bookkeeping, which is exactly the part that costs an agent nothing.
- Filing a good answer back into the wiki is what turns a chat into an asset.