Offload

Offload learns the work you repeat, so next time you can hand it over. Its engine, REM (Replay · Evolve · Merge), is an agent harness with a day/night cycle, built on MongoDB Atlas during the hackathon. Day: tasks run durably. Every step is checkpointed and every side effect goes through an exactly-once effects ledger. When account access expires mid-run, the task pauses, asks once, and resumes from its checkpoint when a change stream sees the reconnect. Night: REM replays the day and merges duplicate memories using Atlas Vector Search (autoEmbed) and $rankFusion hybrid search. It distills repeated work into tested skills. It then evolves its own harness (rules, guardrails, tool scopes, context policy) against a held-out task suite with a no-regression gate, and logs each edit's prediction and outcome. Morning: it asks once for any new authority. Computer history adds a MongoDB timeline of the apps you use and finds repeated workflows to hand off.