Just ran the v0.7.0 update focusing on the new memory module. It's a step forward, but the implementation feels half-baked.
The main changes:
* `Memory` is now a formal module with `save_context` and `get_context`.
* Uses a `VectorStoreRetriever` for recall.
* Defaults to Chroma, which is fine for local testing.
My issue? The abstraction leaks. You still have to manually wire the context saving into the agent's execution loop. Out of the box, it doesn't persist context between tasks automatically. You have to subclass or modify the `execute_task` method. For a "memory module," this feels like a missed opportunity.
Example of the necessary hack:
```python
# Simplified from my test
class MyAgent(BabyAGI):
def execute_task(self, task: Dict):
# ... original logic ...
# Manually save context AFTER execution
self.memory.save_context(
{"input": task["task_name"]},
{"output": result}
)
```
Without this, the memory is just a queryable store. Not truly autonomous.
Also, the Chroma dependency is heavy for a lightweight agent. Wish they'd abstracted the vector store backend more cleanly. Has anyone else integrated a lighter option like Qdrant or just raw FAISS?
Benchmarks or bust.