Bugs found and fixed during end-to-end QA testing of the v3 OSS pipeline
across both the TypeScript and Python SDKs. All changes are OSS-side
only; platform client untouched.
## TypeScript fixes (mem0-ts/)
### 1. Entity extractor trailing punctuation (utils/entity_extraction.ts)
End-of-sentence proper nouns retained their trailing period ("Paris."
vs "Paris"), causing cross-batch entity dedup to fail silently —
embeddings of "Paris." and "Paris" don't hit the 0.95 similarity
threshold. Fix: strip trailing sentence punctuation (. , ; ! ?) in
the existing cleanup pass.
### 2. Undefined filter values leak into vector stores (memory/index.ts)
PR #4843's validateAndTrimEntityId refactor spread agent_id: undefined
and run_id: undefined into every getAll/search filter dict. Qdrant
rejected the malformed match (400), pgvector bound NULL (0 rows),
Redis emitted literal "undefined" in TAG filters. Fix: strip
undefined values via Object.fromEntries filter at both call sites.
### 3. Redis TAG filter values unescaped (vector_stores/redis.ts)
Every UUID contains hyphens, which RediSearch interprets as minus
operators. @user_id:{legacy-abc} parses as "legacy AND NOT abc" —
zero matches. Fix: escapeRedisTagValue helper backslash-escapes all
RediSearch TAG special characters.
### 4. Entity cleanup on delete/update/deleteAll (memory/index.ts)
delete(), update(), and deleteAll() never touched the _entities
collection — linkedMemoryIds accumulated stale ids on every mutation.
Search entity-boost then surfaced deleted or rewritten memories. Fix:
new _removeMemoryFromEntityStore and _linkEntitiesForMemory helpers,
wired into deleteMemory and updateMemory. deleteAll gets coverage for
free (loops deleteMemory).
### 5. textLemmatized missing on infer:false path (memory/index.ts)
createMemory() (used by infer:false) set data and hash but never
called lemmatizeForBm25(). Memories added with infer:false had
degraded BM25 — keyword search fell back to raw data. Fix: one-line
addition of textLemmatized to createMemory payload.
## Python fix (mem0/)
### 6. Entity cleanup on delete/update/deleteAll (memory/main.py)
Same bug as TS #4 — _delete_memory and _update_memory never touched
the entity store. Fix: _remove_memory_from_entity_store and
_link_entities_for_memory helpers (sync + async), wired into both
Memory and AsyncMemory. delete_all covered transitively.
## Verification
All changes verified via scratch QA probes against live vector stores:
TS probes (in-memory SQLite, Qdrant, pgvector, Redis):
- Entity cross-batch dedup: Paris entity merges linkedMemoryIds
- Entity boost at search: +67% score delta with vs without entities
- Delete cleanup: stale id removed, ghost entities deleted
- Update cleanup: old entity unlinked, new entity created
- deleteAll: entity store fully cleared
- Legacy data compat: all 4 stores pass (seed v1 record, v3 read/write)
- BM25 contributes real signal on in-memory store
- infer:false: 0 LLM calls, textLemmatized populated, entities skipped
- No-graph: clean separation, legacy config silently ignored
Python probes (Qdrant server):
- Delete/update/deleteAll entity cleanup: all 3 scenarios pass
TS build: clean (552 unit tests pass)
Python tests: 25 + 35 + 173 pass (2 pre-existing failures unrelated)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Soumil Rathi