prompt(memory): tighten memoryExtractor against transient snapshots + role-name filenames

[RowitZou]

Problem

The 2026-05-20 dogfood report (Bug 3) found memoryExtractor over-producing — one night's memory tree held 113 files. Three symptoms:

1. Transient status snapshots saved as permanent memory — ~24 timestamped status / progress snapshots (stock-monitoring-latest-status-2026-05-19-21-29.md, strength-test-second-round-3-returned.md, …) saved as type:project entries that go stale within the hour.
2. Role names leaked into filenames — webSearcher/generalist-*.md (foreign role prefix), generalist/generalist-*.md (double prefix).
3. -update / -complete filename variants instead of in-place updates.

Root cause is the prompt itself: it explicitly said DO save ... ongoing-work status, directly inducing #1, while the adjacent temporary task context exclusion contradicted it. Filename guidance was only "concise kebab/snake-case" — nothing about provenance or reuse.

Change

Three edits to the memoryExtractor system prompt:

• New ## What counts as durable section with a concrete test — "would this entry be wrong or useless an hour from now?" — and an explicit Skip list for point-in-time status / progress snapshots. Replaces the contradictory "ongoing-work status" / "temporary task context" bullets.
• filename field: name the memory's topic; never prefix with a role / agent / task name; reuse an existing entry's filename verbatim when updating.
• Workflow step 2: when the conversation only refines an existing entry, write to its exact filename — never create a -update / -v2 / -final sibling.

Stays within the prompt design principle: first person, no enumeration of sibling agents / architecture, "the request" framing. Structurally it adds one section and moves the save/skip bullets out of "Body conventions" into it.

Out of scope

The cross-layer duplication symptom (Bug 3 現象2 — main re-extracting <background-task-result> blocks into L1, because main's extraction can't see worker L3 entries) is not a prompt issue. It needs a code change (exclude bg-result blocks from main's extraction window) and is left to a separate PR. Bug 4 having restored the curator also lets it reconcile genuine cross-role duplicates.

Testing

• memoryExtractor role-prompt snapshot hash updated in both snapshot tests (role-prompt.snapshot.test.ts, prompt-for-role.test.ts): 15ad654… → 296e169…. No other role prompt changed.
• The extract.test.ts regression guard ("Do NOT emit memory contents as JSON") still holds — that line is unchanged.
• pnpm typecheck clean; pnpm test 1462/1462 pass.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR updates the bundled memoryExtractor system prompt to reduce low-value memory writes by strengthening “durability” guidance and tightening filename/update rules, and refreshes prompt snapshot hashes to match.

Changes:

• Add an explicit “durability test” + skip list to prevent transient status/progress snapshots from being saved as long-term memory.
• Tighten filename guidance to avoid role/agent/task prefixes and to reuse existing filenames verbatim when updating.
• Update snapshot-hash fixtures for memoryExtractor in the two prompt snapshot tests.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/agents/bundled/memoryExtractor.ts	Updates the memoryExtractor prompt with durability criteria and filename/update rules to reduce transient/duplicate memory files.
src/agents/role-prompt.snapshot.test.ts	Refreshes the expected prompt hash for memoryExtractor.
src/prompt/prompt-for-role.test.ts	Refreshes the expected prompt hash for memoryExtractor.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.