fix(memory): address critical bugs from multi-agent review

Bug Fixes:
- Remove singleton pattern from consolidation/reflection services to
  prevent stale database session bugs (session is now passed per-request)
- Add LRU eviction to MemoryToolService._working dict (max 1000 sessions)
  to prevent unbounded memory growth
- Replace O(n) list.remove() with O(1) OrderedDict.move_to_end() in
  RetrievalCache for better performance under load
- Use deque with maxlen for metrics histograms to prevent unbounded
  memory growth (circular buffer with 10k max samples)
- Use full UUID for checkpoint IDs instead of 8-char prefix to avoid
  collision risk at scale (birthday paradox at ~50k checkpoints)

Test Updates:
- Update checkpoint test to expect 36-char UUID
- Update reflection singleton tests to expect new factory behavior
- Add reset_memory_reflection() no-op for backwards compatibility

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
2026-01-05 18:55:32 +01:00
parent 35aea2d73a
commit 3edce9cd26
8 changed files with 86 additions and 78 deletions

View File

@@ -423,7 +423,8 @@ class WorkingMemory:
Returns:
Checkpoint ID for later restoration
"""
checkpoint_id = str(uuid.uuid4())[:8]
# Use full UUID to avoid collision risk (8 chars has ~50k collision at birthday paradox)
checkpoint_id = str(uuid.uuid4())
checkpoint_key = f"{_CHECKPOINT_PREFIX}{checkpoint_id}"
# Capture all current state