✨ Obliviate Fossilium

"Your memory file remembers things that stopped being true in July."

Memory files accrete like sediment: every correction keeps the old wording as a warning, every migration leaves "note: this changed" markers, every count gets hand-typed and never updated. The model reads all of it, every session, and weighs the fossil as heavily as the living rule.

What to scan

The root memory file, every nested one, whatever the repo names them (CLAUDE.md, AGENTS.md), user-global memory files with the user's consent, and any doc the instruction layer tells agents to read routinely.

Anti-patterns to find

1. Migration-relative phrasing. "This used to say X until <date>", "note this changed", "the old rule was…". Useful for the week after the change; a fossil forever after. The current fact is the only fact the model needs.
2. Retraction narratives in the hot path. Paragraph-length stories of how a claim was wrong and got corrected. The correction earns its context; the story doesn't.
3. Typed facts that drift. Hand-written counts, sizes, version numbers, dates ("the suite has 5,738 tests") that go stale silently. Symptom: two different numbers for the same fact in one file.
4. History as instruction. Vendor-adoption sagas, decision back-stories, who-ruled-what timelines living inline where a decision-record pointer would do.
5. Restated rules. The same rule paraphrased in three sections because each author feared the reader would miss it (also feeds Contradicto Revelio when the copies drift apart).
6. Panic-marker sediment. ⚠ markers on resolved history (→ Quietus Capslockius grades the volume; this pass relocates the content).
7. Sheer size. Official guidance targets lean memory files (on the order of a couple hundred lines); multi-thousand-line roots are paying context tax every single session.

Fix prescriptions

The archive is not a landfill

The extracted history must stay findable: one archive file, dated sections, indexed from wherever the repo indexes knowledge. Deleting history outright destroys the evidence trail that keeps decisions from being re-litigated — move it, don't burn it.

Bounds

• If legilimens-anthropica ran this session, prefer its measured numbers over the ones in this file. This file is a snapshot; that pass is the current reading.
• Scope: memory files — root, nested, and user-global only with the user's consent. Not application source, not docs the instruction layer merely links.
• Cap: 15 findings, ranked by bytes recovered per unit of risk. Class counts carry the rest.
• Done when each memory file has been read once and its fossils classified.
• Quote once, at read time; mark anything you cannot re-confirm UNVERIFIED rather than re-reading to check yourself.
• See skills/opus-pocus/references/casting-safety.md for the full rules.

Report format

Per finding: file:line · class (1–7) · quoted text · destination (rewrite in place / archive / derive). Lead with the before/after: bytes now, projected bytes after, fossil count by class. SCANNED NOTHING if no memory files exist, grading N/A.