Dream

Dream is a batch transform, not an agent: input in, a structured diff out. No loop, no wandering, no per-message tool calls. It runs nightly (and on demand), reads everything gated since the last watermark, and reconciles it into typed state.

memcal dream --dry-run   # preview size and estimated cost; calls nothing
memcal dream             # run the pass

The stages

1. Bundle (code). Everything gated since the last watermark is grouped by entity or thread, across all streams — Jordan's text, the group-chat line where Jordan spoke, and Jordan's email land in one bundle. A typical day is 5–15 bundles. This is the key choice: splitting by source would separate the things that must be joined, while grouping by subject makes cross-platform deduplication structural. Three messages about one dinner arrive in one context, so no model can create three memories from them.

2. Propose (model, parallel, packed). Each call takes up to MEMCAL_PACK_BUNDLES (default 6) bundles and sees their content plus current state — the event window, open to-dos, standing, and the wiki pages for its entities. That state is identical across calls, so it caches. Each returns a typed, keyed diff: event inserts and updates, to-do opens and closes, wiki slot fills, standing edits, questions to ask. Never free-form memories — only diffs against keys, which is what makes deduplication automatic.

3. Merge (code, model only on conflict). Keyed diffs merge deterministically. Two bundles proposing the same row collide on the key and merge. Two bundles touching one wiki page apply per-slot. Only genuinely ambiguous near-horizon cases get a model call.

4. Apply (code). Version-checked writes update the typed stores. Every row records which pass wrote it, and write operations carry the originating message plus the row version they acted on — so re-reading the sentence that moved an event does not create a second one. See Evidence & provenance.

5. Sweep (one cheap model call). Reads the resulting state plus the day's diffs — around 2k tokens, not the day's traffic. Duplicates? Contradictions? Obvious junk? Healing lives here, and it is cheap precisely because it reads output rather than input. Skippable with --no-sweep.

6. Render (code). Writes brief.md, trimmed to the token cap. See Brief.

Write precedence

A correction made during the day is not undone by older evidence collected later. Precedence runs per field on when evidence was said: a message written before a correction cannot revise the corrected field, while genuinely newer evidence still can. Cancelling one booking and making another keeps the old row declined and the new row confirmed; an older notice cannot undo a newer confirmation. Nothing becomes permanently unchangeable.

Dream reviews original source messages and genuine user statements — never the assistant's own prose. A wrong assistant answer stays a wrong answer until real evidence corrects it; reprocessing never promotes it into a correction. Typed corrections applied during the day survive the nightly pass without duplication: the same event identity, the same evidence.

Modes and recovery

One watermark-driven program with --mode nightly|ondemand|realtime (ondemand by default; nightly is the scheduled pass). Short-cadence collection without consolidation runs between passes. --rounds repeats until the spool drains; --redo un-claims processed items for reprocessing (merge-on-key, so corrections are not duplicated); --retry RUN requeues a failed or partial run's readable lines. The Runs view shows whether each pass was ok, partial, failed, running, or priced only. See Dream runs.