What changed
Google released ReasoningBank, an agent-memory framework that turns completed trajectories into structured, reusable strategies. It learns from both successes and failures and combines memory with parallel or sequential test-time exploration; code covers WebArena and SWE-bench experiments.
Why it matters
Useful agent memory should preserve transferable decisions and failure warnings, not replay every raw action or only celebrate successful runs.
What to do
On a sandboxed task set, store one concise lesson per trajectory, require evidence links back to the run, and compare later success and step count with a memory-free baseline.
Caveat
The code is a research demonstration requiring substantial benchmark setup and model access; self-judged memories can encode false lessons.