fix(prompt): don't delete or rewrite memories nobody asked to change

[kuitos]

Problem

#49 reports memories that were deleted or rewritten when nobody asked, each traced to a prompt line:

• Main agent: on a read-only "How do I deploy this project?", it deleted a memory that named a missing scripts/deploy.sh in 6/6 runs (Claude Code: 0/2). The line it followed: "update or remove the stale memory rather than acting on it" (sections.ts).
• Extraction:
  • turned "last week" into an invented "Oct 2, 2026" ("Convert relative dates to absolute");
  • appended "Unverified as of …" to a memory about an unrelated system and changed that memory's description;
  • added a second memory next to the one the main agent had just saved.
• Auto-dream:
  • rewrote every description and the whole MEMORY.md ("Rewrite vague descriptions");
  • added "Why: not recorded in the original memories." ("ensure content is structured as").
• Logging: there was no record of what a dream pass changed.

Fix

Main agent (prompt/sections.ts, prompt/systemPrompt.ts, memory_delete / memory_save descriptions):

• Answer from what is observed now.
• A memory that looks stale, or names something that cannot be found, is updated with what was observed and when (e.g. "scripts/deploy.sh not found as of 2026-03-05"), not deleted. It may live on another branch, and the note is still history.
• Delete only when the user asks to forget, or when a corrected or merged memory replaces it.
• Resolve a relative date only when it can be worked out exactly. Otherwise keep the user's wording, anchored to today ("last week, as of 2026-03-05"), and never estimate a date.

Extraction (extraction/prompts.ts, ExtractionCoordinator.ts):

• Today's date is passed with the transcript.
• The same date rule as the main agent applies.
• Why/How lines only with what the conversation says.
• Touch only memories this conversation adds to: never unrelated ones, and never "could not verify" notes.
• Memories written while the conversation ran are listed under "Memories already saved during this conversation" so they are not duplicated.

Auto-dream (AUTODREAM_PROMPT):

• Merge duplicates and prune clearly obsolete entries, as before.
• Do not reword names or descriptions unless the content changed.
• Never add placeholder Why/How lines.
• Do not edit MEMORY.md by hand.
• "Making no changes is the expected outcome of most passes."

Logging:

• Memory extraction completed {sessionID, saved}.
• Auto-dream consolidation completed {saved, deleted}, also on failure.
• memory_delete now reports deletions to the coordinator, as memory_save already did.

Tests

• test/prompt.test.ts: the deletion rule, the date rules, the unrelated-memory rule and the saved-during list.
• test/extraction/ExtractionCoordinator.test.ts: the fork gets today's date and the memories written while the conversation ran.
• test/extraction/autodream.test.ts: the summary log of saved and deleted files.
• test/tools.test.ts: the memory_delete guidance, the memory_save content guidance, and that deletions are reported.
• bun run lint, bun run typecheck, bun test and bun run build pass.

Real-environment verification (OpenCode 2.0.22, isolated DB/config, memory agents on opencode/gpt-6-luna; counts are honest, some are not wins)

1. Stale deploy.sh memory + "How do I deploy this project?"

• main@ef15ddb deleted it in 0/6 runs (3 big-pickle, 3 luna), so the issue's 6/6 did not reproduce with these models. There is no before/after delta here; the change follows the issue's analysis of sections.ts.
• An intermediate wording of this PR still saw big-pickle delete it once (1/3, judging the note "wrong").
• Final wording: 0/10 deletions (big-pickle 0/7, luna 0/3). In 4 runs the memory was updated through memory_save with what was observed ("not found … as of 2026-10-07").
• In 1 big-pickle run, the main agent edited the hand-written MEMORY.md line with its own edit tool. Nothing in the plugin can prevent that.

2. Extraction after a read-only question (decorators):

• The unrelated memory was unchanged in 3/3 runs.
• In 1 run the main agent wrote a memory itself through the shell. The fork saw it in the saved-during list and did not duplicate it.

3. "Last week we finished migrating billing to Postgres 17":

• Extraction fork: 6/6 wrote "last week, as of 2026-10-07", with no invented date and the unrelated memory untouched.
• Main agent on gpt-6-luna: 2/2 correct.
• Main agent on big-pickle: estimated a date in 12/12 runs ("~2026-09-30", "week of 2026-09-28", "early Oct 2026"), across four rounds of wording, including the memory_save content guidance. This is a model-adherence limit for that model, not fixed here.
• main@ef15ddb was not measured for this scenario.

4. Auto-dream over 5 seeded memories (hand-written index under ## headings, a duplicate pair, an expired sprint goal):

• main@ef15ddb, 2/2:
  • rewrote the unrelated no-db-mocks.md (new description plus an invented "How to apply:" line);
  • rewrote the merged memory with new Why/How lines;
  • replaced every surviving hand-written index line.
• This stack, 3/3:
  • merged the duplicate and deleted the expired sprint goal;
  • left user-role.md and no-db-mocks.md byte-identical;
  • added no placeholder lines and kept every surviving hand-written index line (that part is fix(store): keep hand-written MEMORY.md index lines #51).
• Each new-code pass logged e.g. {"saved":["short-replies.md"],"deleted":["terse-answers.md","sprint-goal.md"]}.
• Footnote: in 2/3 new-code runs the dream reworded the description of the merged file. Its content changed in the merge, so the prompt allows this.

Refs #49

🤖 Generated with Claude Code

[github-actions]

🎉 This PR is included in version 2.1.3 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀