Did Codex Reset
GitHub

Just-in-Time Memory curates raw agent traces when a new task arrives

DAIR.AI

DAIR.AI describes a Salesforce AI Research paper on agent memory called Just-in-Time Memory. Instead of summarizing each run when it ends, the method stores raw trajectories and decides what to extract from them when the next task arrives.

A curator reads the retrieved traces together with the new task and writes a short memory payload for that task. Because the payload is used immediately, the curator can be trained on whether that same task succeeds. DAIR.AI says that even an untrained curator matches or beats memory written at the end of a task.

On ALFWorld, WebShop and tau2-bench, DAIR.AI says Just-in-Time Memory beats the strongest baseline by 16.2, 16.3 and 3.9 success-rate points. The Just-in-Time Memory paper is linked by DAIR.AI.

#just-in-time-memory#salesforce-ai-research#agent-memory#llm-agents

Only 35.8% of the cleanest public terminal-agent RL collection passed a Salesforce AI Research audit

Salesforce AI Research put that collection at 35.8% clean and two others at 10.1% and 3.3%, after finding reward errors in both directions. The authors say RIVER, which filters defective environments, increases RL gains by 106% on Terminal-Bench-Lite and 30% on Terminal-Bench v2.1 while using fewer than 30% of TMax's environments.