openclaw - 💡(How to fix) Fix Feature: openclaw wiki compile - Karpathy-style browsable synthesis from workspace memory [1 comments, 2 participants]

Context

Karpathy LLM Wiki pattern went viral in April: instead of query-time RAG, the LLM compiles incoming sources into a persistent, browsable markdown wiki. 41k+ bookmarks on the gist. Main insight: do the synthesis work once at ingest time, save the output, never re-derive.

OpenClaw workspaces already have all the raw material a wiki-compile needs:

• MEMORY.md long-term facts
• memory/YYYY-MM-DD.md daily logs
• modules/*/CONTEXT.md module state-of-play
• Optionally, an external memory like OpenBrain (separate proposal)

There is no built-in way to compile this into a browsable synthesis today.

Proposal

Add openclaw wiki compile as a first-class command:

• Reads memory files + active module CONTEXT.md files from the workspace.
• Optionally merges in an external memory target.
• Regenerates a wiki/ directory at workspace root with:
  • index.md — catalog of pages with one-line summaries
  • modules/<name>.md — one synthesis page per active module
  • log.md — append-only compile log
• Uses the active OpenClaw model. Safe to run on local Ollama for privacy.
• Idempotent: running on unchanged input produces the same files.

Working reference implementation

I built a working version as a single Node script (~230 LOC, no deps beyond Node fetch). Design notes:

• Deterministic skeleton (file layout, index, log) written by the tool. The LLM only writes the synthesis body per module. Smaller hallucination surface, much cheaper.
• Never touches raw sources. wiki/ is regenerated, not edited.
• Output is genuinely readable and useful — the synthesis pulls recent activity, open loops, and key files from CONTEXT.md cross-referenced against daily memory.

Compile time: 5 modules in ~45 seconds on qwen2.5-coder:7b once the model is warm.

Learned the hard way: Ollama's watchdog kills models at >25% CPU, and gemma4:26b idles at 26-29% during normal generation. Defaulted to qwen2.5-coder:7b which stays comfortably under. Worth documenting either in OpenClaw or in the Ollama watchdog upstream.

Why this fits OpenClaw's ethos

OpenClaw already has the best memory primitives of any agent stack I have used. A compile step turns them into a browsable artifact humans can actually read, without forcing anyone onto an external tool (though it pairs great with Obsidian if you want the graph view). Following the Karpathy pattern keeps OpenClaw on the right side of the "do work once, not every query" divide.

What I would like from maintainers

1. Is this in scope for core, or should it stay a user-space tool?
2. If in scope: draft PR, or start with a design doc?

— Simon (running the full OpenClaw stack on Windows 11 home setup with local Ollama)