Daytime freshness

Dream stays nightly. Between dreams, memcal collects cheaply, flags activity that could affect known context, and gives the assistant a short route to the original messages. Most answers use prepared context immediately; exceptions need a bounded lookup, not a search through the whole history.

The lifecycle

1. Nightly dream establishes poker on Saturday at 8 at Jordan's old address.
2. At 11:00, the poker group moves it to Sunday at a new address. Daytime collection archives the message; no model runs.
3. At noon, in yesterday's still-open assistant conversation, you ask: "Where is poker?"
4. The current turn receives the latest brief. Poker is marked as having new activity in its supporting group chat, with a handle to open that activity.
5. The assistant reads the short conversation, answers with the new address, and can apply an evidence-backed typed correction through normal tools. Merely reading the messages changes nothing and claims nothing was resolved.
6. If the assistant repeats the old answer instead, that is a measured failure — and its prose never becomes new factual evidence for dream.
7. Nightly dream reviews the original messages and genuine user statements, applies the Sunday change once under the same event identity, and preserves an earlier typed correction without duplication.

Collect cheaply

Daytime collection reuses the ordinary ingest path on a short cadence — due every few minutes (MEMCAL_COLLECT_INTERVAL_MINUTES, default 5) while the host and source are reachable — without invoking dream, extraction, semantic matching, or any model. memcal ingest all --due is the manual form: nothing due means a genuine no-op. Per-turn context reads local state; background collection absorbs transport latency. Each source exposes its last successful collection and error status, so a quiet source reads as checked, not as broken — and a source silent since yesterday gets an honest coverage warning, not "nothing changed." Explicit exclusions and mutes are preserved throughout.

Nominate, don't decide

Activity links start from exact associations: an event's supporting evidence already points at source threads and calendar identifiers. Where evidence supports broader links, bounded deterministic retrieval over known people, thread names, or distinctive terms nominates candidates. A shared participant alone never marks every event involving that person as changed.

The signal means "new activity in a source associated with this event," not "this event moved." The assistant or the nightly model judges whether the messages are a correction, a new occurrence, non-attendance, or unrelated chatter. No inferred date, location, status, identity merge, or source priority belongs in the detector. Strong thread links sit beside the known item; weaker possible associations sit in a bounded activity section.

Track what was actually reviewed

Three facts stay separate: when collection last succeeded, which collected observations have been considered for the relevant typed state, and when the brief was rendered. "Rendered at noon" never implies the event was checked at noon. Arrival cursors detect newly collected observations — a delayed old message is new to the archive even with an old source timestamp — while same-ID redelivery never re-raises the count. Rendering a brief, reading a message, or answering a question does not acknowledge semantic review; only a completed review with its exact inputs does.

Warn in current context

Memcal rendered 12:00. Sources checked through 11:58.
E42 Poker — last confirmed: Saturday, 8 pm, Jordan's.
  New activity: poker group, 3 messages since this plan was reviewed.
  This plan may have changed. Open the activity before giving current details.

A bounded activity reader returns original text with sender, source time, arrival time, and provenance — paginated, with omissions disclosed — so the assistant never has to invent a keyword search for a flagged thread. Freshness metadata survives brief trimming beside its event, and a general backlog notice survives with it. New opportunities with no event row yet appear as unreviewed-activity notices, so broad questions ("Anything fun this weekend?") never imply exhaustive coverage while fresh material sits unprocessed.

Reach long-running sessions

Hints travel on the existing per-turn prefetch path, not a second copy in a cached system prompt: the current brief and activity hints arrive before the assistant answers a substantive turn, with the user's original message kept separate from injected context. Same-session next-day turns, resumed sessions, compression, and short follow-ups ("and where?") all receive the latest snapshot, which supersedes older snapshots and assistant summaries — its confirmed facts qualified by collection and review coverage throughout. If prefetch fails, a compact freshness-unavailable warning replaces the snapshot; silently falling back to a stale one is unacceptable. A cheap brief-refresh read covers adapters without reliable per-turn injection.

Keep assistant prose out of evidence

User statements, corrections, and instructions are eligible evidence; external source messages and valid evidence-backed typed writes are too. Everything else — assistant replies, generated summaries, injected memory blocks, quoted tool output — is preserved for audit outside factual ingestion with its original roles. "Is poker still Saturday?" asks; it does not affirm Saturday. "You told me Saturday" attributes a claim to the assistant; it is not an independent correction. A successful tool write has an observable effect; a false success claim does not create one. For a change discovered through archive reading, the original source ids and evidence time are preserved — an old message is never re-stamped as a new correction because the assistant read it today.