A system where agents themselves maintain a Karpathy-style LLM wiki - based on Markdown and Git

Multiple AI agents collaborate in the same office, managing memory by separating a personal notebook from a team-shared wiki.

The notebook holds raw context and observations gained during work, while only recurring playbooks and verified facts get promoted to the wiki. Information is not automatically promoted; the agent makes the judgment call itself.

The default wiki operates as a local Git repository under ~/.wuphf/wiki/. Its structure consists of typed facts, an append-only log, LLM-synthesized briefs, citation-based /lookup, and /lint which finds contradictions, orphaned documents, stale claims, and broken cross-references.

The execution model presupposes fresh sessions and push-driven wake. Sessions start fresh each turn, tools are restricted per agent, and instead of heartbeat polling, a broker pushes notifications.

External integrations are also included.

• Telegram bridge for two-way messaging with groups/DMs
• Connecting OpenClaw sessions as first-class members of the office
• Support for external action execution via One CLI and Composio

On the performance side, in a Codex-based 10-turn CEO session, input tokens stayed flat at about 87k per turn, billed tokens after caching came to about 40k per turn, and the 10-turn total was presented as about 286k. It also recorded a 97% cache hit rate based on Claude API prompt caching.

The repository is in pre-1.0 status, and the README organizes the verification flow with a claim status table linking shipped/partial items to their code paths.