Sleep-time compute
Memory that organizes itself while your agents are idle.
Without memory
Naive memory systems do all their thinking at recall time — every query pays for disorganized storage. Raw captures pile up, duplicates accumulate, and the graph fills with stale edges nobody prunes.
With Rekall
Rekall does its housekeeping off-peak, the way brains do: background workers consolidate short-term into long-term memory, infer new graph edges, decay stale ones, and age preferences — so recall hits an organized store, not a junk drawer.
How it works
Capture stays cheap
Ingest is fast and raw; nothing blocks on organization at write time.
Consolidation runs while you sleep
Workers dedupe, summarize, and promote significant short-term memories into long-term storage on a schedule.
The graph tends itself
Inference adds edges you never stated (A relates to C via B); decay prunes what stopped being true or useful.
See it in action
03:00 [ltm-consolidation] 212 STM episodes reviewed→ 34 promoted to LTM, 9 merged as duplicates03:10 [graph-inference] 12 new edges inferred(co-occurrence + transitive)03:20 [graph-decay] 41 stale edges downweighted,3 invalidated facts archived (bi-temporal)03:25 [preference-aging] 6 preferences decayedRecall tomorrow hits a cleaner store than today.