Sleep-Time Compute

Memory work moved off the turn — consolidation, summarisation, reflection — run between conversations rather than while someone is waiting.

Why it matters in a memory layer

It is cheap and it is not free. On this course's corpus, deferring every consolidation to a single batch costs 38 embeddings against 503, and reaches the identical store — consolidation is order-independent here, which is the property that makes deferral safe at all and is worth measuring rather than assuming.

What it costs is the window. Deferred, the store spends 11 of 24 turns believing a job the user has already left, including the turn where they correct the assistant about it. Indexing late means a fact is missing; arbitrating late means the system argues.

Connections

Taught in: Sleep-Time Compute

Used in: Background Job Mechanics · The Latency Budget

Do not confuse with: Deduplication