Memory Also Learns

This is an implementation that turns agent memory from a simple vector DB wrapper into a local memory engine that learns based on usage outcomes.

The core idea is to assign a score to each memory, reinforcing it with reinforce_if_used when the agent succeeds using that memory, and penalizing it when it fails. Over time, this allows the system to distinguish which "lessons" are valid and which context was noise.

The implementation features are as follows.

• Fully local execution: works without OpenAI or the cloud
• ONNX embedding: uses MiniLM-L6-v2
• Performance: about 23MB in size, ~6ms recall at 1k memories
• Integration: supports Python / Node / Go via a Rust shared library
• MCP server added: can be connected directly to Claude Desktop, Cursor

Through the README's "Float Money Bug" demo, it shows that simple similarity search alone cannot sufficiently guarantee the memory quality of a code agent.