What does a knowledge base look like when its main reader is an AI?
The LLM Wiki
Thrivbe's LLM Wiki is an Obsidian vault the AI both reads and writes, using three retrieval layers — a 500-word hot page, an index catalog, then the pages themselves — with no vector search until the vault earns it. Nightly jobs compile work sessions and meeting transcripts into it, so knowledge compounds instead of being rediscovered.
- Obsidian
- Claude
- Nightly compile jobs
Relevant services: Thrivbe AI
Hypothesis
Every AI session rediscovers the same context from scratch. If an AI maintains its own wiki — ingesting sources, compiling what it learned each day, linking entities across business, research, and personal domains — knowledge should compound instead of evaporating.
What we built
An Obsidian vault the AI reads and writes. Three retrieval layers: a ~500-word "hot" page of recent context, an index catalog, then the pages themselves — deliberately no vector search until the vault earns it. Skills for ingesting any source into 5–15 page updates, querying with citations, and linting vault health. A nightly job compiles the day's work sessions into the vault; another ingests meeting transcripts — 492 historical meetings were backfilled. Conflicting facts get an explicit conflict callout rather than a silent overwrite.
Learnings
- Layered retrieval (hot page → index → pages) covers most queries without any RAG infrastructure. Structure first, embeddings later.
- The compounding effect is real but only if ingestion is a habit — which is why it runs on a timer, not on remembering.
- "Never silently resolve a conflict" turned out to be the most important convention: an AI that quietly overwrites disagreeing facts poisons its own well.
Log
- 2026-07-11 — Vault now git-backed with nightly offsite backup and a secret-scan gate on every push.
- 2026-06-06 — 492 meeting transcripts backfilled; nightly backlog ingest armed.
- 2026-06-05 — Vault designed and seeded.