Advanced · Background processing
Sleep-Time Compute: Consolidating Memory in the Background
Sleep-time compute processes and consolidates agent memory between sessions — like biological sleep — producing cheaper, smarter recall on the next interaction without blocking the user-facing turn.
Sleep window
Definition
What is sleep-time compute?
Async batch processing while the agent is idle — summarize, merge, extract facts, re-embed and prune without blocking the user-facing turn.
Also called background consolidation. Trending term in 2025–2026 agent research: move expensive memory work off the hot path to between-session windows, like biological sleep consolidating short-term into long-term memory.
Benefits
Benefits of sleep-time compute
- Lower latency — user turns skip consolidation LLM calls
- Better LTM — batch merge produces cleaner long-term facts
- Cost shift — run expensive jobs off-peak or in background workers
- Higher recall quality — more time for extraction and deduplication
Engram’s async write pipeline is sleep-time compute in practice — fire-and-forget memory builds while the app continues (Weaviate, June 2026).
Implementation
Implementation patterns
- Session-end job queue — trigger consolidation when user disconnects
- Cron batch — nightly merge/summarize for all active users
- Event-driven workers — queue memory jobs off the hot path
- Letta archival compaction — background paging and merge in archival tier
What runs in the sleep window: summarize transcript → extract facts → merge duplicates → re-embed if needed → evict raw transcript.
FAQ
Frequently asked questions
Sleep-time compute vs realtime consolidation?
Realtime consolidates during the user turn (higher latency). Sleep-time batches consolidation between sessions (lower hot-path latency, delayed LTM update).
Production examples of sleep-time compute?
Engram's async write pipeline (Weaviate GA June 2026), session-end summarization jobs, Letta archival compaction, nightly Mem0 batch merges.
Does Mem0 support sleep-time consolidation?
Engram runs fire-and-forget async extraction pipelines natively. Mem0 writes asynchronously in managed API; for explicit batch consolidation, trigger merge jobs on session end.
Letta archival memory and sleep-time?
Letta compacts and pages archival memory in background — MemGPT pattern. DMR 93.4% (Packer et al., 2023). See Letta alternatives.
Cost savings from sleep-time compute?
Shifts LLM calls off hot path; batch embedding jobs at off-peak rates. Mem0 ~1,800 vs ~26,000 tokens per query regardless of when consolidation runs (Chhikara et al., 2025).
Sleep-time compute vs summarization?
Summarization is one sleep-time task. Full sleep window also runs merge, re-embed, prune and fact extraction. See memory summarization.