[Phototonic]

Summary

When the memory plugin injects remembered knowledge into a session, those memories used to be anonymous: the agent had no way to say which memory it was relying on, how old it was, or where it came from. This change gives every injected memory a visible identity — its id, creation date, and project — and tells the agent it may cite that id when leaning on a memory, so "I remember that we chose libSQL" can become "I remember that we chose libSQL (mem_1792183254076_wxsig8s9y, recorded 2026-10-16)". The same provenance is attached to the memories restored into a session after compaction, and the ids match what the Memory Explorer web UI shows, so a cited memory can be found and inspected later.

Details:

• Provenance in the injected memory block, so agents can cite what they "remember":

• Each <memory> tag in the injected <memory_context> block now carries id, created (UTC YYYY-MM-DD), and project attributes when present. Rows without provenance keep the exact previous tag shape.

• The block header tells agents they may cite a memory by its id when relying on it.

• Compaction's ## Restored Session Memory headings gain the same provenance: ### Memory N (mem_… · created YYYY-MM-DD · project), suffix omitted entirely when absent.

• Shared helpers live in a new pure module (src/services/memory-provenance.ts). This keeps the mocked context.js export surface intact — several scenario tests mock.module that file with a minimal surface, so src/index.ts must not import new exports from it.

Session-title lookup was intentionally descoped per the issue (project name stands in).

Test plan

• Exercised the change locally — a live session quoted its injected block verbatim: <memory relevance="100%" id="mem_…" created="2026-10-20" project="opencode-mem"> plus the new header sentence (the issue's acceptance criterion)
• Noted any follow-up or known limitations — exact-format unit tests for both blocks in tests/memory-provenance.test.ts; three low-severity follow-ups from the pre-push review round are recorded in a comment below

Checklist

• Branch is based on main (rebased onto 6289ac4; single commit)
• bun test passes (787 pass / 3 fail — all three pre-existing on Linux main: plugin-bundle-boundary, user-profile-learning-confidence, config storagePath shape)
• bun run typecheck passes
• bun run check passes (format + lint + typecheck)
• Docs / README updated — N/A: the injected block format is not documented in README, so no doc surface changes

[Phototonic]

Pre-push review round (independent read-only reviewer over the full diff): APPROVE WITH COMMENTS — no blockers. Findings and disposition:

1. Low — formatMemoryCreatedDate accepts timezone-free strings (2026-10-20T00:30:00 parses relative to local TZ; 2026-02-30 normalizes to Mar 2). Not reachable from current callers — they supply epoch-ms or Z-suffixed ISO (safeToISOString / raw created_at). Follow-up candidate: treat unzoned input as UTC + focused tests.
2. Low — the compaction suffix renders id/projectName verbatim, so a project name containing embedded newlines could disturb the Markdown heading. IDs are plugin-generated (no newlines) and memory bodies already enter these prompts verbatim, so this is formatting hardening only. Follow-up candidate: collapse control whitespace in display metadata.
3. Low — header wording states every memory carries an id/date while legacy rows intentionally omit them. Kept as originally proposed; happy to soften to "when available…" if maintainers prefer.

Reviewer validation: 46 targeted tests across six files + typecheck passed. Full suite rerun before push: 787 pass / 3 pre-existing Linux main failures (plugin-bundle-boundary, user-profile-learning-confidence, config storagePath shape).

Supersedes #421 (same change, renamed branch).