Problem

Closes #2648

The metadata dict passed by callers to _create_memory and _create_procedural_memory (both sync and async) was mutated in-place. These methods directly set keys like data, hash, created_at, and memory_type on the caller's original dict object.

This causes two problems:

1. Concurrency/thread-safety: When the same metadata dict is shared across concurrent calls (common in async usage with asyncio.gather), one call's mutations bleed into another, producing corrupted payloads.
2. Surprising side-effects: Even in sequential usage, the caller's dict is silently modified after the call returns, which violates the principle of least surprise.

Affected methods

Note: _update_memory and _add_to_vector_store already use deepcopy and are not affected.

Solution

Use deepcopy(metadata) at the entry point of each affected method before any mutation, storing the result in a new_metadata local variable. The caller's original dict is never touched.

# Before (mutates caller's dict)
metadata = metadata or {}
metadata["data"] = data

# After (safe copy)
new_metadata = deepcopy(metadata) if metadata is not None else {}
new_metadata["data"] = data

This is a minimal, backward-compatible change — no caller reads back from the metadata dict after passing it to these methods.

Testing

Added 12 new tests in TestMetadataNotMutated covering:

• No-mutation guarantee: Verify the caller's metadata dict is identical before and after calling _create_memory, _update_memory, and _add_to_vector_store (sync + async)
• Shared dict safety: Pass the same metadata dict to multiple sequential _create_memory calls and verify it remains unchanged, while each call stores the correct distinct payload
• Nested structure protection: Verify deepcopy protects nested dicts and lists inside metadata
• Payload correctness: Verify the stored payload still contains all expected fields (data, hash, created_at, role, actor_id, custom fields)
• History correctness: Verify role and actor_id from metadata flow through correctly to add_history after deepcopy
• None metadata: Verify metadata=None still works correctly
• End-to-end no-infer path: Verify _add_to_vector_store(infer=False) with multiple messages doesn't leak metadata between iterations (sync + async)

All 109 existing + new tests pass.

🤖 Generated with Claude Code