The dreaming pass

The pass fixes two named limits at once. In-band memory curation splits an agent's attention between the task and the bookkeeping, so both suffer. And any one session can only see itself — a failure pattern recurring across a fleet is invisible to every individual agent, because each context window starts empty. The fix is a second-order process with its own token budget: fan sub-agents across a transcript batch, judge which patterns are prevalent, propose individual edits — memory entries, skill changes, guardrails — each carrying provenance and prevalence. A human accepts or rejects each one. Nothing lands unreviewed.

The wider framing: when the model's weights are closed or too slow to change, the surfaces you own are the harness, the context, the memory and the skill library — so improvement has to accrue there as an explicit engineering loop. At production scale the loop has a characteristic economics: cluster traces cheaply first, discard the intended-behaviour majority, and spend frontier-model analysis only on the anomalous tail.

The distillation half has a discipline of its own. The output is not a summary of the transcript but a belief that cites the span which produced it, so memory stays auditable and revisable instead of accreted prose. And the strongest sources argue the durable unit is the distilled skill, not the raw experience — raw-trace memory starts to hurt once the model can absorb procedure.

The consolidation pass — out-of-band transcript mining proposing evidence-cited updates through human review — is Assay's largest standing design debt in the memory theme. What this scan changed is its status: with a production implementation at Replit and a production shape at Anthropic, it is a build item, not a research bet.

Proposals land as per-entry record files through a pull request, never as direct memory writes — which is how the registers already work, so the concurrency and review machinery exists. Every proposal carries at least one transcript citation plus a prevalence count; an uncited proposal is rejected by construction, not by reviewer vigilance. The pass ingests more than chat — tool calls, guard refusals, red-CI events, review comments, human interruptions — because the strongest worked examples turn on signals that never appear in conversation. Cheap clustering runs before expensive diagnosis. Git is the concurrency layer: a rejected stale push is compare-and-swap. And the pass proposes skill edits at least as often as memory facts — skills are the vehicle for distilled procedure, memory files for facts and rulings.

A minimum bar for the first version: humans accept at least half of what the pass proposes. Below that, the gate is filtering machine noise rather than reviewing insight, and the pass is not paying for its tokens.

• Whether anyone reports an accept rate for machine-proposed memory or skill edits — the number that says whether the human gate is a formality or a filter.
• Whether the pass fires on a cron or on an activity threshold in the systems that ship it.
• Whether the cluster-then-diagnose ordering appears in a second production system, or stays Replit's.
• Whether anyone prices the pass against the improvement it produces — every source describes the mechanism; none costs it.