llm-wiki

Institutional memory for AI agents — and the agent team that uses it.

Great PMs don't remember everything; they run a process: write decisions down, record who dissented, brief the team before the meeting, review the work. Your codebase has git — but your product knowledge (decisions, research, rejected options, "what would change our mind") usually lives in scattered docs and heads, where neither humans nor AI agents can reliably find it.

llm-wiki is that process, packaged as a Claude Code plugin, in two halves:

• The wiki (the memory): a persistent markdown knowledge base — decision pages with lifecycle fields, trust-graded claims ([verified] / [hypothesis] / [REFUTED]), research digests, session logs. Based on the Karpathy LLM Wiki pattern: knowledge compounds instead of being re-discovered every session. Every wiki is an Obsidian vault — you browse your team's knowledge in a friendly app, no terminal required.
• The factory (the organization): persona teams that read that wiki with budgeted context, sign their own positions (agents are never paraphrased), close out sessions idempotently — minutes, decisions, action items — and propose improvements to their own instructions that only you can approve.

You bring the vision, the problem statements, and the judgment on results. The system runs the team — and remembers everything.