bug fix: unexpected vector store reset when Memory.delete_all()

[peisong13]

Fix: Memory.delete_all() incorrectly wiping entire vector store collection

Problem Description

The Memory.delete_all() method had a critical bug where it was calling self.vector_store.reset() after deleting filtered memories. This caused the entire vector store collection to be wiped, not just the memories that matched the provided filters (user_id, agent_id, run_id).

Impact

• Data Loss: When users called delete_all(user_id="user1"), it would delete user1's memories as expected, but then wipe ALL memories from ALL users
• Unexpected Behavior: The method name suggests filtering, but the implementation caused a complete reset
• Production Risk: This could cause catastrophic data loss in production environments

Root Cause

In mem0/memory/main.py, line 1053, the delete_all() method contained:

# delete vector memories by filter
memories = self.vector_store.list(filters=filters)[0]
for memory in memories:
    self._delete_memory(memory.id)
# self.vector_store.reset()  # <-- This line was the problem

The self.vector_store.reset() call was incorrectly resetting the entire vector store collection after the filtered deletion.

Solution

Simple Fix: Comment out the problematic self.vector_store.reset() line.

Why This Fix Is Correct

1. Filtered Deletion Already Works: The code above the reset call correctly:

   • Queries for memories matching the filters
   • Deletes each memory individually by ID
   • This preserves unfiltered memories as intended

2. Reset Is Unnecessary: The reset() method is designed to clear ALL data, which contradicts the filtering behavior users expect from delete_all(user_id=...). According to the docs, delete_all() should only delete filtered memories (https://docs.mem0.ai/core-concepts/memory-operations/delete). Resetting the whole vector store is unexpected behavior.

3. Consistent with Method Contract: The method signature and documentation indicate filtered deletion, not a complete reset

Changes Made

• File: mem0/memory/main.py
• Line: 1053
• Change: Commented out # self.vector_store.reset()

How Has This Been Tested?

• Test Script

def test_delete_all_does_not_call_vector_store_reset(mock_memory_instance):
    """
    Test that delete_all() only deletes specific memories by ID and does not call vector_store.reset().
    
    This test verifies the bug fix where delete_all() was incorrectly calling
    self.vector_store.reset() which wiped the entire collection.
    """
    # Setup: Create mock memories that would be returned by list()
    mock_mem1 = MockMemory("mem_1", {"data": "User likes pizza", "user_id": "user1"})
    mock_mem2 = MockMemory("mem_2", {"data": "User likes pasta", "user_id": "user1"}) 
    mock_mem3 = MockMemory("mem_3", {"data": "Agent learned something", "user_id": "user2"})
    
    # Mock vector_store.list to return only user1's memories
    mock_memory_instance.vector_store.list.return_value = ([mock_mem1, mock_mem2], None)
    
    # Execute: Call delete_all with user_id filter
    with patch('mem0.memory.main.process_telemetry_filters') as mock_telemetry:
        mock_telemetry.return_value = (["user_id"], ["encoded_user1"])
        with patch('mem0.memory.main.capture_event') as mock_capture:
            result = mock_memory_instance.delete_all(user_id="user1")
    
    # Verify: Check that only specific memories were deleted by ID
    assert mock_memory_instance._delete_memory.call_count == 2
    mock_memory_instance._delete_memory.assert_any_call("mem_1")
    mock_memory_instance._delete_memory.assert_any_call("mem_2")
    
    # Critical: Verify that vector_store.reset() was NOT called
    mock_memory_instance.vector_store.reset.assert_not_called()
    
    # Verify the list was called with correct filters
    mock_memory_instance.vector_store.list.assert_called_once_with(filters={"user_id": "user1"})
    
    # Verify return message
    assert result["message"] == "Memories deleted successfully!"

def test_delete_all_preserves_unfiltered_memories():
    """
    Integration test to verify that delete_all() only removes filtered memories
    and leaves other memories intact. This simulates the real-world scenario
    where the bug would cause data loss.
    """
    # This test would ideally use a real vector store, but we'll simulate it with mocks
    with patch('mem0.utils.factory.EmbedderFactory.create'):
        with patch('mem0.utils.factory.VectorStoreFactory.create') as mock_vector_factory:
            with patch('mem0.utils.factory.LlmFactory.create'):
                with patch('mem0.memory.storage.SQLiteManager'):
                    
                    # Setup mock vector store
                    mock_vector_store = MagicMock()
                    mock_vector_factory.return_value = mock_vector_store
                    
                    # Simulate vector store state with memories for different users
                    user1_memories = [
                        MockMemory("mem_1", {"data": "User1 memory 1", "user_id": "user1"}),
                        MockMemory("mem_2", {"data": "User1 memory 2", "user_id": "user1"})
                    ]
                    user2_memories = [
                        MockMemory("mem_3", {"data": "User2 memory 1", "user_id": "user2"}),
                        MockMemory("mem_4", {"data": "User2 memory 2", "user_id": "user2"})
                    ]
                    
                    # Mock list method to return only user1 memories when filtered
                    def mock_list(filters=None):
                        if filters and filters.get("user_id") == "user1":
                            return (user1_memories, None)
                        elif filters and filters.get("user_id") == "user2":
                            return (user2_memories, None)
                        else:
                            # Return all memories (this simulates what reset would affect)
                            return (user1_memories + user2_memories, None)
                    
                    mock_vector_store.list.side_effect = mock_list
                    
                    # Track deletions
                    deleted_ids = []
                    def mock_delete(vector_id):
                        deleted_ids.append(vector_id)
                    
                    mock_vector_store.delete.side_effect = mock_delete
                    
                    # Create memory instance
                    from mem0.memory.main import Memory as MemoryClass
                    memory = MemoryClass()
                    
                    # Mock _delete_memory to track what gets deleted
                    memory._delete_memory = MagicMock(side_effect=lambda mem_id: deleted_ids.append(mem_id))
                    
                    # Execute: Delete all memories for user1
                    with patch('mem0.memory.main.process_telemetry_filters') as mock_telemetry:
                        mock_telemetry.return_value = (["user_id"], ["encoded"])
                        with patch('mem0.memory.main.capture_event'):
                            result = memory.delete_all(user_id="user1")
                    
                    # Verify: Only user1's memories were deleted
                    assert "mem_1" in deleted_ids
                    assert "mem_2" in deleted_ids
                    assert "mem_3" not in deleted_ids  # user2's memories should be preserved
                    assert "mem_4" not in deleted_ids  # user2's memories should be preserved
                    
                    # Verify reset was not called (critical for the bug fix)
                    mock_vector_store.reset.assert_not_called()
                    
                    # Verify list was called with proper filter, not without filter (which reset would imply)
                    mock_vector_store.list.assert_called_with(filters={"user_id": "user1"})
                    
                    assert result["message"] == "Memories deleted successfully!"

Test Results

Before/After Behavior

Before (Buggy)

memory.add("User1 likes pizza", user_id="user1")
memory.add("User2 likes pasta", user_id="user2") 
memory.delete_all(user_id="user1")
# Result: ALL memories deleted (user1 AND user2)

After (Fixed)

memory.add("User1 likes pizza", user_id="user1")
memory.add("User2 likes pasta", user_id="user2")
memory.delete_all(user_id="user1")
# Result: Only user1's memory deleted, user2's memory preserved

Related Issues

This fix resolves the data loss issue where filtered deletions would unexpectedly wipe the entire memory collection.

[CLAassistant]

All committers have signed the CLA.

[peisong13]

related issue: #3746