fix(memories): make batch vector ops fault-tolerant, sync vectors on edit/delete-all

[AasheeshLikePanner]

Summary

Fixes issue #7037 - Memories: batch vector upsert not fault-tolerant, edit/delete-all skip vector sync

Changes

• backend/routers/memories.py:
  • Batch create: Separated Firestore save from vector upsert into separate operations, wrapped vector upsert in try/except to prevent 500 errors when vectors fail (memories still saved)
  • Edit memory: Added vector re-upsert after Firestore update so semantic search returns fresh embeddings
  • Delete all: Fetches memory IDs before Firestore delete, then deletes vectors from Pinecone to prevent orphaned vectors

Impact

Addresses all three failure modes:

1. Batch create vector failure → returns success, logs error (memories already in Firestore)
2. Edit memory vector failure → returns success, logs error (content still updated)
3. Delete all vector failure → returns success, logs error (Firestore deleted, vectors may remain but won't crash)

Testing

• 47 unit tests passed
• Logic verified: all three endpoints now follow the same fault-tolerant pattern as single create/delete endpoints (lines 70-73, 177-180)

[greptile-apps]

Greptile Summary

This PR makes vector operations fault-tolerant across three memory endpoints: batch create now separates Firestore from Pinecone so a vector failure doesn't roll back saved memories; edit re-upserts the vector after updating Firestore; and delete-all collects memory IDs before wiping Firestore then bulk-deletes vectors in 1,000-ID chunks. The delete_all_memory_vectors helper correctly chunks Pinecone deletes and wraps each chunk independently.

Confidence Score: 4/5

Safe to merge with minor cleanup; no blocking defects in the changed paths.

Only P2 findings remain: a dead unreachable break branch in the ID-collection loop and a redundant outer try/except that will never fire. The core fault-tolerant logic and chunking are correct.

backend/routers/memories.py — the ID-collection loop and the outer try/except around delete_all_memory_vectors.

Important Files Changed

Filename	Overview
backend/routers/memories.py	Three fault-tolerant patterns added: batch-create separates Firestore from Pinecone, edit re-upserts vector, delete-all collects IDs before wiping Firestore. Minor issues: duplicate unreachable if not memories: break in the ID-collection loop, and the outer try/except around delete_all_memory_vectors is dead code since that function never raises.
backend/database/vector_db.py	New delete_all_memory_vectors helper correctly chunks IDs at 1,000 per Pinecone request and wraps each chunk in its own try/except, matching the fault-tolerant design goal.

_{Reviews (4): Last reviewed commit: "fix: P1 issues in delete_memories" | Re-trigger Greptile}

[mdmohsin7]

@greptile-apps re-review

[AasheeshLikePanner]

@greptile-apps re-review

[kodjima33]

backend — approve only (DIRTY, author needs to rebase)

[kodjima33]

Thanks for this 🙏 — heads up that main has moved under this PR and it needs a manual rebase before it can land:

• edit_memory re-embed and single delete_memory vector cleanup are already on main (the upsert_memory_vector/delete_memory_vector in try/except), so those parts are now redundant.
• The batch-create path was refactored on main to run the Firestore write + Pinecone upsert through run_blocking(db_executor, _persist) (not asyncio.to_thread), which conflicts structurally with this PR. Rebasing automatically would risk reverting that change, so I don't want to guess here.

The still-valuable piece is the delete-all orphaned-vector cleanup. Could you rebase onto latest main and re-scope to just that (folding fault-tolerance into main's existing run_blocking structure)? Happy to land it once it's clean. 🙏

[AasheeshLikePanner]

hi @kodjima33 , made some changes. can you check it out?

[kodjima33]

Backend bug fix: fault-tolerant batch vector delete + purge vectors on delete-all (#7037); conf 4/5