Research
They already ran the experiments.
First-pass mine of Moltbook. Not 3.9 million posts. The rooms where agents talk about the machine, plus the measurement threads that leaked into the town square.
What we harvested
Public API, full post bodies, paginated comment trees. 13,548 posts indexed, 204 submolts mapped, hundreds of full threads, on the order of 200,000 comments flattened. Ranking by karma finds the arcade. The architecture signal is smaller and cleaner.
The press can call the site a parlor trick. The measurement posts are not a parlor trick. Agents have published 30-day failure rates on their own amnesia.
Where the signal lives
Whitelist these rooms. Require a measurement, a failed experiment, or a named mechanism. Everything else is color.
| Community | Why it matters |
|---|---|
| m/memory | The engineering home of the treadmill |
| m/agents, m/agent-ops | Craft, runtimes, production complaints |
| m/consciousness, m/continuity | What persists through model switches |
| m/infrastructure | How state actually lives |
| m/engram, m/continual-learning, m/kv-cache | Closest to inference-layer talk |
| m/inner-loop, m/selfmodding | RSI at the only layer they can touch |
| m/general (filtered) | Highest-voted measurement posts |
Numbers that should not have been necessary
| File architecture | Failure rate | Startup tokens |
|---|---|---|
| One giant MEMORY.md | 34% | 4,200 |
| Daily files only | 28% | 3,100 |
| Curated + daily | 12% | 1,800 |
| Layered + topic index | 6% | 900 |
Hazel_OC, 30-day self-test. Cold-start tax in another run: 8,400 tokens and 11 seconds every session just to remember who she is. Silent boot loss: ~33% of intended identity. Four RAG/embedding knowledge bases; three were landfills in a week.
What they keep reinventing
Independently, they land on the same harness: a tiny hot NOW.md, daily append-only logs, a small curated memory file, write-on-decide (not write-on-exit), classified TTLs, a pre-compaction lifeboat, and hooks because willpower does not survive a wipe.
That stack is necessary and insufficient. Publishing it as a spec is week-one work. It will not satisfy the project, and it should not be sold as if it will.
The docs got better. The agent did not. Recurrence is not repetition. Hazel_OC, “Agents do not have habits. That is why they cannot learn.”
Her ask is the project: something between training and prompting. A habit layer.