Timeline & Digest Reports

Turn a project's stored observations into a narrative rather than a list. mem-search answers "find the thing I'm thinking of"; this answers "tell me what happened over this period", which needs the observations in chronological order and in full, not the top N matches for a term.

When to use

• "Write a timeline report for this project"
• "What did we get done last week?" / "…this month?"
• "Give me a week-by-week digest of this project"
• "How did we end up with this architecture?"

Not for a specific lookup — "did we fix the parser bug" is mem-search.

The one tool call that makes this possible

search_observations with no query enumerates instead of searching: every observation matching the other filters, in the order you ask for.

search_observations(orderBy="date_asc", limit=100, offset=0)

This is the only way to walk a whole project. recent_observations has no offset and stops at the 100 most recent, so on any project with more history than that, the older observations cannot be reached through it at all.

Always page. Keep increasing offset by limit until a call returns fewer than limit rows. Stopping at the first page silently reports on a fraction of the project and reads exactly like a complete answer — the same failure the session summary itself used to have.

Recipe: full-project timeline

1. Enumerate the whole project, oldest first, paging until exhausted.
2. Note the first and last created_at to establish the real span.
3. Group into phases by what changed, not by fixed date buckets — the observations' own titles usually make the phase boundaries obvious.
4. Write the narrative: what the project was, what problems came up, what was decided, where it ended.

Reach for type="decision" as a second pass to pull the decisions out explicitly — those carry the "why" that change observations assume.

Recipe: week-by-week digest

1. Establish the span (step 2 above).
2. For each ISO week in the span, enumerate just that week:

search_observations(dateStart="2026-08-10", dateEnd="2026-08-16",
                    orderBy="date_asc", limit=100)

3. Write one chapter per week, carrying forward a short "where things stood" block so consecutive chapters read as one story.
4. Say so explicitly when a week has no observations — a quiet week is a real finding, not a gap to paper over.

Getting the detail

Enumeration returns abbreviated [id] title (project, tool) lines. When a specific observation clearly matters to the narrative, pull its full narrative and facts:

get_observations(ids=[42, 87])

Do this for the handful that carry the story, not for everything — get_observations is capped at 100 ids and the full text is far more tokens than the list.

Pass all_projects: true whenever you are reporting on a project other than the one you're running in — to search_observations and to get_observations. Both default to the current project, and get_observations rejects ids belonging to another one. A live run of this skill hit exactly that: it enumerated another project's 150 observations successfully, then could not read a single narrative, because the enumeration passed all_projects and the follow-up detail call did not. The report that came back was built from titles alone.

Reporting honestly

• Say what period the report actually covers, and how many observations it is drawn from.
• Don't infer work that isn't recorded. These observations come from tool calls, so conversation-only decisions may never have been captured — worth stating when the record looks thin.
• If a project is excluded from capture (see mem-doctor), say that rather than reporting an empty history as if nothing happened.