NUS, Princeton, Stanford & Oxford unveil Recuris: recursive self-improving agent memory without training

jiqizhixin · x · 2026-09-07

A team from NUS, Princeton, Stanford, and Oxford introduced Recuris (Recursive Experiential–Working Memory Evolution), targeting a fundamental flaw in agent memory for long-horizon tasks: memory updates driven only by downstream task scores can't pinpoint which component caused a failure, yielding coarse, untargeted updates that degrade further as reasoning traces grow.

Recuris is claimed to be the first agent architecture to apply recursive self-improvement inside the memory control layer. It maintains an experiential–working memory that continuously aligns accumulated experience with current execution needs, using three core technical breakthroughs to identify and fix specific memory components — with no training required.

The authors report it works across models from 3B open-source models up to Claude Opus 5 and GPT-5.6-Sol (post truncated; detailed results in the original).

Original post →

More from coding & agent

coding & agent channel →