🐛 Describe the bug

When Memory._add_to_vector_store() in main.py determines a memory is unchanged (NONE event, ~line 578), it calls:

self.vector_store.update(
    vector_id=memory_id,
    vector=None,        # no new vector, content didn't change
    payload=updated_metadata,
)

In valkey.py, the update() method unconditionally serializes the vector:
"embedding": np.array(vector, dtype=np.float32).tobytes(),

When vector is None, np.array(None, dtype=np.float32) produces a single float32 scalar (4 bytes), which overwrites the correct embedding (e.g. 4096 bytes for a 1024-dim model). The memory's text content remains intact, but it becomes permanently unsearchable via vector similarity because the embedding dimensions no longer match the index schema.

Steps to reproduce:

1. A memory is stored (correct embedding written)
2. The same fact is encountered again in a subsequent add() call
3. The LLM returns a NONE action (content unchanged)
4. update() is called with vector=None to update session metadata
5. The embedding is overwritten with 4 bytes of garbage