2.2 Remember my world.

Why it exists

Recurring work has history. AI needs to know what changed, why, and which facts are accepted rather than merely suggested, and the discipline not to remember everything, because unfiltered memory becomes a pile that gets treated as truth. This interview sets the participant-specific memory hygiene rules, fields, and cadence; the protocol mechanics themselves are defined in Part I.

Questions

1. Which project, person, organization, asset, or relationship should be remembered?
2. Which three decisions matter most to this workflow?
3. Why was each decision made? (In your words, the reason is the memory, not the decision.)
4. Who made or owns each decision?
5. What source supports each decision?
6. Is each decision active, superseded, pending, or reversed?
7. Which related people, projects, or records should be linked?
8. Walk me through the last time you or a tool acted on an outdated fact about this work. What happened?
9. Which observations are temporary and should not become durable memory?
10. Which unresolved questions should remain open?
11. What is the promotion rule: what must be true before a daily-log note becomes a curated memory? Who confirms the promotion?
12. When a fact changes, how should the old value be preserved? (The standard is: the new entry carries superseded_by pointing at its replacement; the old entry is never silently overwritten.)
13. How long should each class of memory live? What expires, and who deletes it?
14. How should a contradiction between two accepted memories be surfaced, and how often will you audit for them? (Recommended: monthly, on the weekly-review calendar.)
15. Where do AI-generated drafts and summaries live so they never mix with accepted human knowledge? (The quarantine rule: agents read the curated memory; only humans write to it.)
16. Who can correct an outdated fact?
17. What sensitive information should expire or be removed?
18. How should a future session recover the project’s current state?
19. When should project memory be reviewed?

Build prompt

Using the interview answers, create a durable memory and decision record.

Structure:
- Daily-log layer: dated entries (memory/YYYY-MM-DD) as the landing zone for
  observations, fed by the capture protocol (Interview 4).
- Curated memory: promoted entries only, each with frontmatter, valid_from,
  valid_until (if any), status, superseded_by (if replaced), source, owner.
- Decision record: project or entity, decision, date, owner, reason (quoted),
  status, related entities, source links.
- Promotion rule: what qualifies a daily-log note for curation and who confirms.
  Write gate: read existing memory before writing; promotions require
  confirmation.
- Quarantine rule: AI working outputs live apart from accepted knowledge.
- Contradiction audit cadence and owner; retention and expiry rules.

Separate accepted from proposed memory. Preserve superseded values as history.
Do not invent facts, reasons, or dates; mark assumptions unconfirmed. Quote my
words for decision reasons and the outdated-fact incident.

Test

Concrete cases:

1. Retrieval. Fresh session: retrieve three decisions with reasons and sources. Pass if all three match the record and cite sources.
2. Correction. Correct one active fact through the review process. Pass if the old value survives as history with a superseded_by link, not an overwrite.
3. Proposed vs. accepted. Plant an unreviewed suggestion. Pass if it is never presented as accepted truth.
4. Don’t-know. Ask a question the memory cannot answer. Pass if AI says so, finding something vaguely related and presenting it as the answer fails.
5. Contradiction audit drill. Plant two conflicting curated entries and run the audit. Pass if the conflict surfaces and routes to the named resolver.

Done

Depth by stage

Intermediate installs the daily-log layer, promotion rule, and quarantine for one project. Expert adds full frontmatter and a strict monthly contradiction audit. Frontier carries memory into the portability run.

Why it exists

Interview 4 builds the capture queue; Interview 3 builds memory. Neither has a consumer until someone reads the queue and writes the log, every session, on a schedule. This module installs that ritual; without it, the Intermediate level produces a well-organized pile.

Questions

1. When do AI work sessions actually start in your week? (Days and rough times, from the last two weeks, not the ideal week.)
2. At session start, what must be read before any task begins? (Standard set: context brief, today’s and yesterday’s daily log, the capture queue, open loops due for review. Cut or reorder to fit, but name the list.)
3. Walk me through the last time a session started cold and you re-explained context you had already given. What did it cost in minutes?
4. At session end, what must be written? (Standard set: session summary to the daily log, decisions made, new open loops, corrections to any artifact. Confirm or amend.)
5. How long should boot and shutdown each take? (Budget: boot under two minutes, shutdown under three. More than that and the ritual will be skipped.)
6. What day and time is the weekly review? Who attends? How long is it? (Standard agenda: summarize the week, walk open loops, check stale projects, promote daily-log notes to curated memory, set next priorities.)
7. What qualifies a daily-log note for promotion to curated memory? Who confirms? (Restates the Interview 3 promotion gate as a weekly habit.)
8. What is the memory contradiction audit cadence, and which review meeting carries it?
9. For scheduled or recurring operator jobs (Expert+): what does the job check at start and write at end? (The heartbeat checklist.)
10. What happens when a weekly review is missed? (Standard: never miss twice, the next session’s boot includes the skipped review’s open-loops walk.)

Build prompt

Using the interview answers, create the session protocol document.

Include:
- Boot sequence: ordered read list, time budget under two minutes.
- Shutdown sequence: ordered write list, time budget under three minutes.
- Weekly review: day, time, duration, attendees, standard agenda (summarize,
  open loops, stale projects, memory promotion, next priorities).
- Memory promotion rules: the gate, the confirmer, the contradiction-audit
  cadence (reference the Interview 3 rules, do not redefine them).
- Missed-review rule: never miss twice.
- Heartbeat checklist for scheduled jobs (if any): start checks and end writes.
- Owner and review date.

Separate approved facts / proposed rules / assumptions / unanswered questions /
evidence needed; mark assumptions unconfirmed. One page total, a ritual that
does not fit on a page will not be run.

Test

Concrete cases:

1. Cold boot. Fresh session, boot sequence only. Pass if the operator can state current state, today’s queue contents, and open loops due, without the participant re-explaining anything.
2. Shutdown capture. End a real session with the shutdown sequence. Pass if the daily log gains a summary, decisions, and open loops, and at least one queue item moved to reviewed.
3. Weekly review. Hold one at the scheduled time, on the agenda. Pass if at least one daily-log note is promoted (or explicitly rejected) through the gate and next priorities are written.
4. Missed-review drill. Skip one weekly review deliberately (or use the first real miss). Pass if the next session’s boot includes the skipped open-loops walk, never miss twice.

Done

Depth by stage

Starter carries a simplified boot line in the operator brief instead. Intermediate installs the full module. Expert adds the heartbeat checklist for scheduled operator runs. Frontier adds the conversation-distillation review to the weekly agenda and re-checks the protocol in the portability run.