Did Codex Reset
GitHub

Jev-Mem paper reports faster memory construction for AI agents

elvis

The paper Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents proposes an agentic memory architecture inspired by System-One and System-Two cognition. A lightweight controller assigns memory types and relations when storing information. During reading, it handles routing, the retrieval budget, graph traversal, scoring, and when to stop. The LLM is called only for final reasoning and answer writing.

On LoCoMo, the paper reports an overall score of 0.777 with an LLM judge, an 11.0% relative improvement over the strongest baseline. It also reports building agent memory 6.6x faster and cutting query latency by 36.7%. Memory construction takes 158 seconds, and average query latency falls to 0.93 seconds.

#jev-mem#ai-agents#agent-memory#locomo

Best Taste-Bench model chooses the better direction 59.7% of the time

A paper from Microsoft and colleagues reports that the best model answers 59.7% of Taste-Bench questions correctly when choosing the better open direction in long engineering and research tasks. Forks decided by later evidence are much harder, a larger reasoning budget does not raise accuracy, and distilling a teacher’s outcome judgment improves held-out SWE-bench Pro success.