🔮 The Remembrall

"It glows when you've forgotten something. You forgot to write it down."

Every other spell in this book reads the instruction layer and reasons about it. This one reads what actually happened. When a human corrects an agent, that correction is ground truth about a missing or broken instruction — and no amount of re-reading the memory file will produce it, because the file's whole defect is that the thing isn't in it. Corrections are the cheapest evidence a repo already has and never looks at.

The book's own doctrine says verify the artifact, never the call. The artifact here is the transcript.

What to scan

Session transcripts, in place. On Claude Code these are JSONL under ~/.claude/projects/<url-encoded-cwd>/*.jsonl — the path is undocumented and version-specific, so confirm it before promising anything:

ls ~/.claude/projects/*/ 2>/dev/null | head

If the store is absent, unreadable, or holds fewer than a handful of sessions, this pass reports ? — could not verify. That is a distinct verdict from a clean one: no transcripts means nothing was examined, not that nothing is wrong.

Alongside them, the resolved memory files and rule files — not to audit their prose, but to answer one question per finding: is this correction already written down somewhere?

Privacy. Transcripts are the user's own local data and often contain far more than this pass needs. Read them in place, quote the minimum that makes a finding checkable, and transmit nothing anywhere. If a correction cannot be quoted without dragging in unrelated content, describe it and cite the line rather than pasting it.

Anti-patterns to find

1. Repeated correction. The human said substantially the same thing three or more times across sessions, and it appears in no instruction file. The rule already exists — it lives in the human's fingers instead of the repo. Its wording is already dictated by the corrections.
2. Ignored instruction. A rule exists and the transcripts show the agent violating it anyway. This is the highest-value class and the most often misdiagnosed: the fix is almost never "state it more forcefully". Diagnose why first — the rule is unfindable (in a file nothing loads at the moment it matters), out-scoped (its trigger never matches the real situation), or outranked (a louder rule nearby wins every collision).
3. Never-triggered instruction. A rule whose trigger situation has not occurred once in recorded history. Untriggered text is untested text. Do not delete it here — hand it to avada-verbosa, which names this candidate class and has no way to mine it.
4. Workflow fiction. Instructions describing a process the transcripts show nobody follows — a documented route that is always taken differently, a ritual described in the file and absent from every session. Either the workflow moved and the file didn't, or the workflow was aspirational from the start.
5. Correction cluster with no rule. A topic the human keeps steering, in different words each time, with no single repeated phrase to match on. Harder to detect than class 1 and worth more: it is a missing rule whose scope the human has been drawing by hand, one session at a time.
6. Correction reversal. The human corrected the agent toward X in one session and toward not-X later. Either the intent changed and the file records the old one, or two situations look alike and need distinguishing. Never propose a rule from one half of a reversal without saying the other half exists.

Fix prescriptions

Bounds

• Scope: the transcript store for this repo, plus the resolved memory and rule files for the "is it written down?" check. Not other projects' transcripts, not application source, not another installed plugin's artifacts.
• Cap: at most 15 findings, ranked by how often the correction recurred. Sample rather than exhaust — the most recent 30–50 sessions answer this well, and reading a year of history to find one more instance is the unbounded sweep this book bounds.
• Done when the sampled window has been read once and the clusters ranked. Not when corrections stop appearing; on a real project they never do.
• Quote once, at read time, and minimally; mark anything you cannot re-confirm UNVERIFIED rather than re-reading to check yourself.
• Read-only, and local-only: no transcript content leaves the machine.
• See skills/opus-pocus/references/casting-safety.md for the full rules.

Report format

Lead with the denominator: sessions available, sessions read, date range. Then per finding: class (1–6) · the correction quoted minimally with its session and line · how many times it recurred · whether an instruction already covers it (and where) · the prescription. Rank by recurrence count, not by how interesting the topic is.

? — could not verify when no transcript store was found or it was unreadable — never a clean grade. SCANNED NOTHING if the store exists but held no sessions for this repo.