llm-wiki

A persistent, LLM-maintained knowledge base built from agent session transcripts (Claude Code and Codex) plus optional Slack and Granola activity. Modeled on Karpathy's LLM Wiki idea: knowledge compounds in markdown rather than being rediscovered every session.

Layers

• Raw sources (immutable, outside this repo): Claude Code transcripts in ~/.claude/projects/**/*.jsonl, Codex rollouts indexed by ~/.codex/state_5.sqlite, optional Slack history, and optional Granola notes.
• The wiki (wiki/): LLM-generated markdown organized by entities, concepts, and syntheses. Owned end-to-end by the system.
• The schema (SCHEMA.md): the rules the LLM follows when ingesting, updating, querying, and linting the wiki.

Navigation

• wiki/entities/projects/ — one page per repo / service / product
• wiki/entities/people/ — collaborators, reviewers, stakeholders
• wiki/entities/technologies/ — libraries, frameworks, tools
• wiki/concepts/ — recurring patterns, design decisions, durable problems
• wiki/syntheses/ — cross-cutting analyses (open questions, decisions log, recurring bugs, lint reports)
• wiki/index/ — auto-derived (by-project, by-date, glossary)
• worklog/ — the transient layer: in-flight workstreams (status, next action, blockers) and their archive. Separate from the durable wiki; see worklog/WORKLOG.md

How it stays alive

Two agent-session ingest paths feed the same pending queue: