fix: clean up graph store data on Memory.delete() by utkarsh240799 · Pull Request #4505 · mem0ai/mem0

utkarsh240799 · 2026-03-23T12:48:04Z

Linked Issue

Description

When deleting a memory via Memory.delete(memory_id), the graph store (Neo4j, Memgraph, Kuzu, Neptune, Apache AGE) was not cleaned up — only the vector store and history DB were updated. This led to orphaned graph nodes and relationships accumulating over time, causing data inconsistency, storage bloat, and stale query results.

Root cause: _delete_memory() only handled vector store deletion + history recording. No graph backend had a delete() method for single-memory cleanup.

Fix: Add a delete(data, filters) method to every graph backend that reuses the existing entity extraction pipeline (_retrieve_nodes_from_data → _establish_nodes_relations_from_data → _delete_entities) to identify and remove graph relationships associated with the deleted memory text. Hook this into Memory.delete() and AsyncMemory.delete().

Key design decisions

No new Cypher queries — reuses each backend's existing _delete_entities() method, preserving per-backend semantics (Neo4j soft-deletes via r.valid=false, all others hard-delete via DELETE r)
No double vector store fetch — delete() fetches the memory once and passes it to _delete_memory() via a new optional existing_memory parameter (backward-compatible default None)
Resilient — graph cleanup is wrapped in try/except at both the Memory level and graph backend level, so failures never block vector store deletion
delete_all() unchanged — continues to use bulk graph.delete_all(), no per-memory graph cleanup (avoids N×LLM calls)
No interference with add() pipeline — _delete_memory() (called during implicit LLM-driven deletes in _add_to_vector_store) does NOT do graph cleanup, because the graph pipeline (_add_to_graph) runs in parallel and manages its own conflict resolution independently

Files changed

File	Change
`mem0/memory/graph_memory.py`	Add `delete(data, filters)` — Neo4j, soft-delete
`mem0/memory/memgraph_memory.py`	Add `delete(data, filters)` — Memgraph, hard-delete
`mem0/memory/kuzu_memory.py`	Add `delete(data, filters)` — Kuzu, hard-delete
`mem0/graphs/neptune/base.py`	Add `delete(data, filters)` — Neptune (both DB and Analytics), hard-delete
`mem0/memory/apache_age_memory.py`	Add `delete(data, filters)` — Apache AGE, hard-delete
`mem0/memory/main.py`	`Memory.delete()` and `AsyncMemory.delete()` now perform graph cleanup; `_delete_memory()` accepts optional pre-fetched memory

Type of Change

Bug fix (non-breaking change that fixes an issue)
New feature (non-breaking change that adds functionality)
Breaking change (fix or feature that would cause existing functionality to change)
Refactor (no functional changes)
Documentation update

Breaking Changes

N/A

Test Coverage

I added/updated unit tests
I added/updated integration tests
I tested manually (describe below)
No tests needed (explain why)

Unit tests (`tests/test_graph_delete.py` — 15 tests, always run in CI)

Test	What it verifies
`test_delete_calls_graph_cleanup_when_graph_enabled`	`graph.delete()` called with correct text and filters
`test_delete_skips_graph_when_not_enabled`	Zero behavior change when `enable_graph=False`
`test_delete_continues_if_graph_cleanup_fails`	Graph exception doesn't block vector deletion
`test_delete_skips_graph_when_no_user_id`	No graph call when memory lacks `user_id`
`test_delete_skips_graph_when_no_memory_text`	No graph call when memory has no text data
`test_delete_passes_all_filters_to_graph`	`user_id`, `agent_id`, `run_id` all forwarded
`test_async_delete_calls_graph_cleanup`	Async variant works correctly
`test_async_delete_continues_if_graph_cleanup_fails`	Async resilience on graph failure
`test_delete_raises_for_nonexistent_memory_with_graph_enabled`	`ValueError` raised, graph untouched
`test_async_delete_raises_for_nonexistent_memory_with_graph_enabled`	Async variant of above
`test_delete_all_does_not_trigger_per_memory_graph_cleanup`	`delete_all()` uses bulk `graph.delete_all()` only
`test_internal_delete_memory_does_not_trigger_graph_cleanup`	`_delete_memory()` does NOT call `graph.delete()` (safe for parallel `add()` pipeline)
`test_graph_memory_delete_calls_internal_methods`	Neo4j `MemoryGraph.delete()` calls the correct internal pipeline
`test_graph_memory_delete_skips_when_no_entities`	No-op when LLM extracts zero entities
`test_graph_memory_delete_handles_exception`	Backend-level exception caught and logged

Kuzu e2e tests (`tests/test_graph_delete_e2e.py` — 14 tests, skipped if `kuzu` not installed)

Real Kuzu embedded database with mocked LLM/embedder:

Test	What it verifies
`test_add_creates_nodes_and_edges`	Baseline: `add()` creates real graph data
`test_delete_removes_edges_created_by_add`	`delete()` removes relationships from real DB
`test_delete_only_removes_matching_edges`	Selective: only targeted relationships removed
`test_delete_with_different_user_id`	User isolation at graph level
`test_delete_nonexistent_relationship_is_safe`	No-op on missing data
`test_delete_with_llm_failure_does_not_raise`	LLM crash handled gracefully
`test_delete_with_empty_entity_extraction`	Empty LLM response → graph untouched
`test_delete_all_removes_everything_for_user`	Bulk delete baseline
`test_add_delete_add_cycle`	Add → delete → re-add works
`test_memory_delete_triggers_graph_cleanup`	Full `Memory.delete()` stack cleans graph
`test_memory_delete_with_graph_preserves_other_users_data`	User isolation at Memory API level
`test_memory_delete_graph_failure_still_deletes_vector`	Graph failure → vector deletion proceeds
`test_memory_delete_all_uses_bulk_not_per_memory`	`delete_all()` uses `graph.delete_all()`
`test_memory_delete_nonexistent_raises_without_graph_side_effects`	Non-existent → `ValueError`, graph untouched

Docker e2e tests (`tests/test_graph_delete_docker.py` — 23 tests, skipped if Docker DBs unavailable)

Tests against real database instances with mocked LLM/embedder. All auto-skip in CI.

Backend	Docker image	Tests	Delete semantics verified
Neo4j	`neo4j:5.23`	6	Soft-delete (`r.valid=false`) confirmed
Memgraph	`memgraph/memgraph`	6	Hard-delete (`DELETE r`) confirmed
Apache AGE	`apache/age`	6	Hard-delete (`DELETE r`) confirmed
Neptune	Neo4j as OpenCypher proxy	5	Hard-delete + `user_id` string type assertion

Each backend tests: add, delete, selective delete, user isolation, delete_all, add-delete-add cycle.

Manual testing

All 23 Docker tests were run locally against real Neo4j 5.23, Memgraph, and Apache AGE containers (all passed). Neptune was tested via Neo4j since it uses the same OpenCypher query language.

Checklist

My code follows the project's style guidelines
I have performed a self-review of my code
I have added tests that prove my fix/feature works
New and existing tests pass locally
I have updated documentation if needed

🤖 Generated with Claude Code

When deleting a memory via Memory.delete(), the graph store (Neo4j, Memgraph, Kuzu, Neptune, Apache AGE) was not cleaned up — only the vector store and history DB were. This led to orphaned graph nodes and relationships accumulating over time. Add a delete(data, filters) method to every graph backend that reuses the existing entity extraction pipeline (_retrieve_nodes_from_data → _establish_nodes_relations_from_data → _delete_entities) to identify and remove graph relationships associated with the deleted memory. Hook this into Memory.delete() and AsyncMemory.delete(), with try/except so graph failures never block vector store deletion. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

kartik-mem0 · 2026-03-23T13:38:41Z

hey @utkarsh240799 please address the failing ci

…assertion - Docker e2e tests now do a fast TCP socket check before attempting database connections, so they skip instantly in CI where no Docker containers are running (previously the driver-level timeout was too slow or the connection returned a Mock) - Fix test_main.py::test_delete to expect the new existing_memory second argument passed to _delete_memory() Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-authored-by: utkarsh240799 <utkarsh240799@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

utkarsh240799 requested review from kartik-mem0 and whysosaket March 23, 2026 13:37

whysosaket approved these changes Mar 23, 2026

View reviewed changes

whysosaket merged commit 5332741 into main Mar 23, 2026
7 checks passed

whysosaket deleted the fix/graph-cleanup-on-memory-delete branch March 23, 2026 13:57

utkarsh240799 mentioned this pull request Mar 23, 2026

Fix/neo4j cleanup on delete 3245 #3557

Closed

12 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: clean up graph store data on Memory.delete()#4505

fix: clean up graph store data on Memory.delete()#4505
whysosaket merged 2 commits intomainfrom
fix/graph-cleanup-on-memory-delete

utkarsh240799 commented Mar 23, 2026

Uh oh!

kartik-mem0 commented Mar 23, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

utkarsh240799 commented Mar 23, 2026

Linked Issue

Description

Key design decisions

Files changed

Type of Change

Breaking Changes

Test Coverage

Unit tests (tests/test_graph_delete.py — 15 tests, always run in CI)

Kuzu e2e tests (tests/test_graph_delete_e2e.py — 14 tests, skipped if kuzu not installed)

Docker e2e tests (tests/test_graph_delete_docker.py — 23 tests, skipped if Docker DBs unavailable)

Manual testing

Checklist

Uh oh!

kartik-mem0 commented Mar 23, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Unit tests (`tests/test_graph_delete.py` — 15 tests, always run in CI)

Kuzu e2e tests (`tests/test_graph_delete_e2e.py` — 14 tests, skipped if `kuzu` not installed)

Docker e2e tests (`tests/test_graph_delete_docker.py` — 23 tests, skipped if Docker DBs unavailable)