AI agents that work over long stretches usually just pile up notes as they go. This new method lets them go back and fix the notes that turned out to be wrong.
A paper posted to arXiv describes ReHarness, a framework for letting language model agents revise their own memory while running, rather than only appending to it. After each step, the system looks at what actually happened and uses that outcome to rewrite the reflections it pulled from memory for that step, pushing the correction backward through its stored state. It does this without extra replay or validation data, working strictly on the live stream of interactions. The researchers tested it across three continual learning benchmarks and three different language model backbones, and it beat baseline approaches using raw context, standard memory and reflection systems, and prompt optimization.
The real finding here is about where today's agent memory breaks down. Most systems treat memory as append-only: once a bad generalization gets written down, it sits there poisoning every future decision until something forces a rewrite. ReHarness builds the correction step into the loop itself, which matters for anything meant to run for days or weeks without a human cleaning up its assumptions.
It is one arXiv submission with results on the authors' own benchmarks, not a deployed system, so the claims about stability and plasticity are promising but self-reported for now.