Wave 1 — Compaction prompt improvement:
- Add three wording-reuse bullets to buildCompactionPrompt() under
CRITICAL MEMORY RULES: do not create rephrased duplicates, reuse
existing wording exactly when re-emitting, only emit new memories
when the fact is new, materially corrected, or more specific.
- This attacks the root cause of zero reinforcement: compaction
generating variant text for the same durable fact.
Wave 2 — Bug fixes:
- Bug #2: Add placeholder comment to superseded_existing branch in
decision dedupe (unreachable until v1.5.4 numbered refs). Preserve
as const type assertions.
- Bug #3: Add memory_migration_superseded evidence event type. Both
P0 and quality cleanup migrations now produce evidence events for
superseded entries. loadWorkspaceMemory appends migration evidence
on first-load migrations only (idempotent via migration IDs). No
historical backfill.
- Bug #4: Add documentation comment explaining that feedback identity
key returns exact key (absorbed_identity currently impossible for
feedback). Add test verifying this behavior.
Wave 3 — Validation baseline script:
- Add scripts/dev/validate-identity-keys.ts: read-only script that
scans workspace memory stores, computes exact/identity key
collisions, and reports reinforcement statistics. Baseline matches
audit: 0 exact collisions, 0 identity collisions, 0 reinforcement
events across 123 active memories.
Identity extension is gated on measurement: if the prompt change
produces measurable reinforcement (reinforcementCount > 0), identity
extension may be unnecessary. Decision dedupe stays exact-only
(Wave 4 deferred).
Implement OQ-2 decision: allow at most one reinforcement per memory
identity per UTC calendar day. Same-day reinforcement is blocked
regardless of session or interval. This prevents repetitive-task
gaming where a daily recurring task could reach MAX_COUNT=6 in hours.
Guard order: same-session → calendar-day → 1-hour → max-count
(existing guards kept as defense-in-depth)
1 hour guard is redundant within same day but preserved for
sub-hour edge cases.
Move retention constants and math to a focused src/retention.ts module:
- All half-life, reinforcement, dormancy constants
- TYPE_FACTOR, SOURCE_FACTOR, USER_IMPORTANCE_FACTOR
- RETENTION_TYPE_MAX (renamed from TYPE_MAX)
- calculateInitialStrength, calculateEffectiveHalfLife,
calculateRetentionStrength, calculateDormantDays,
calculateEffectiveAgeDays, reinforceMemory
No behavior changes. retention.ts imports only types from types.ts.
Workspace-memory.ts still owns storage, consolidation, and rendering.
- Remove SAFETY_CRITICAL_FACTOR = 6.0 from workspace-memory.ts
- Remove safetyFactor from calculateInitialStrength() - all memories now
fade according to the same rules
- Remove safetyCritical bypass from applyTypeMaxCaps() - safetyCritical
entries compete normally under TYPE_MAX caps
- Preserve safetyCritical?: boolean in LongTermMemoryEntry type for
backward compatibility (no producer sets it to true)
- Update memory-diag to show deprecation warning instead of capacity alert
- Update tests: add backward-compatibility fixture test, deprecation
strength test, normal cap competition test
- Update docs/architecture.md, RELEASE_NOTES.md, CHANGELOG.md,
docs/configuration.md
Phase 1.5 complete: safetyCritical is now a deprecated field with no
active behavior. Safety rules belong in user-controlled agent.md files.
P0.1 - Fix dormant effective age formula:
- Use overlap logic: only apply dormancy to entry's lifetime
- Formula: activeDays + dormantOverlapDays * 0.25
- calculateDormantDays now returns total days (not excess past grace)
- Test: 28 dormant days → 17.5 effective days
P0.2 - Remove hard stale pruning:
- Remove isPrunableByAge from enforcement
- Remove rejected_stale from accounting reasons
- Elimination now by cap competition only
P0.3 - Integrate reinforcement:
- Call reinforceMemory in dedupe absorption path
- Call reinforceMemory in promotion duplicate path
- Update retentionClock on reinforcement
A1 - Retention clock reset on reinforcement
A4 - Fix tests to encode correct formula
Wave 2c - Dormant workspace tracking:
- Add lastActivityAt to WorkspaceMemoryStore
- Implement calculateDormantDays with 14-day grace period
- Wire dormant days into retention-strength calculation
Wave 3 - Reinforcement:
- Add lastReinforcedSessionID to LongTermMemoryEntry
- Implement reinforceMemory with guards (same-session, 1hr interval, max 6)
- Set retentionClock on memory creation in extractors.ts and plugin.ts
Tests: 219 → 222, all pass
Problem: clearPendingMemories() and recordPromotionRejections() would
incorrectly clear or mutate owned entries during global unowned promotion.
Fixes:
1. clearPendingMemories() now respects owner/unowned scope:
- global clearUnowned only clears unowned same-key entries
- owned same-key entries are preserved
- explicit global clear-all-by-key fallback still works
2. recordPromotionRejections() now has includeUnownedOnly option:
- global unowned rejection only increments/exhausts unowned entries
- owned same-key entries are preserved
3. Added regression tests:
- global unowned clear keeps owned same-key entries
- global unowned rejection only exhausts unowned same-key entries
Tests: 182 pass, 0 fail
Problem: CI test "updateJSON serializes writes across separate node processes"
was failing with expect 100 but got 89/97. The root cause was isLockStale()
being too aggressive - it could mistakenly delete locks held by other processes.
Fixes:
1. isLockStale() now uses mtime only - fresh locks are never stale
2. Added heartbeat mechanism during lock hold to support long updaters
3. Removed PID check that was unreliable in CI/containers
4. Fixed ENOENT race when lock is released between EEXIST and stat
Tests: 180 pass, 0 fail
Wave 3 of memory quality optimization plan.
- Add good memory examples in buildCompactionPrompt()
- Add bad memory examples to skip (test counts, commit hashes, etc.)
- Add prompt assertions in tests to prevent regression
- Emphasize 'useful if a new agent opens this workspace next week'
Architecture review fixes:
- Promotion accounting: only clear pending memories that survived
workspace memory normalization/cap limits. Use retainedKeys from
the returned normalized store instead of attemptedKeys.
- Shared sessionID extraction: add sessionIDFromEventProperties()
helper and use it in both session.compacted and session.deleted,
fixing the previous gap where session.deleted only read info.id.
- Strengthen compaction refresh test: seed workspace memory before
first transform so firstSystem1 is non-empty, then assert
refreshed system[1] preserves existing entries AND contains
promoted memories.
P0c fixes:
- Chinese file count regex now accepts 個/个 between number and 文件
- Admin PIN short reference (<20 chars) passes via config value allowlist
- Phase snapshot uses semantic window (.{0,20}) instead of absolute position
P0d fixes:
- Feedback key split: 500 error and port issue remain separate entries
- extractEntityKey avoids over-merging unrelated plugin configs
- chooseBetterMemory supports supersession mode (newer beats longer)
- Sort comparator now includes source priority as secondary tie-breaker
New regression tests (11 total):
- Real Admin PIN short reference passes
- Real Chinese 37 個文件 snapshot rejected
- Real pathology Phase 1-4 snapshot rejected
- Feedback 500 vs port entries not collapsed
- Unrelated plugin configs not collapsed
- Supersession prefers newer shorter over older longer
67/67 tests pass.
P0a: Parser now accepts both - [type] text and - type text formats
P0b: Prompt adds durable-content guidance to avoid session-specific snapshots
P0c: Parser quality gate rejects exact test counts, file counts, phase progress
- Only rejects phase progress when it appears early in the string (snapshot)
- Stable config values with numbers (Admin PIN, Scrypt) still pass
- Adds 7 new tests covering bracketless parsing and snapshot rejection
Root cause: OpenCode's default compaction template uses --- separators.
When our plugin adds structured context (Memory candidates: format), the
model strictly follows the template, outputting --- at position 0. The
markdown textmate grammar treats this as YAML frontmatter, applying the
'comment' syntax scope (purple + italic in themes like palenight).
Fix: Set output.prompt in the compacting hook to replace the entire
template with a ---free version. Uses only ## Markdown headings and
explicitly forbids YAML frontmatter, horizontal rules, and delimiter
lines. Preserves context from other plugins by merging output.context.
- Replace compactionContextHeader() with buildCompactionPrompt()
- Set output.prompt instead of pushing to output.context
- Merge existing output.context from other plugins before clearing
- Add 'Instructions' section to the template (per architect review)
- Update tests: verify output.prompt, ---free format, context merging
Ralph Chang