ReHarness Rewrites Agent Memory Midstream

ReHarness targets a real weakness in persistent agent memory: stale advice can remain useful-looking while repeatedly causing errors. Its backward revision loop is a compelling step toward memory that learns from consequences, though it adds inference cost and depends heavily on the quality of retrieved reflections.

• –Revises memories after use instead of merely appending new reflections
• –Corrects overgeneralized guidance under distribution shifts
• –Achieves the strongest average normalized gain among evaluated baselines
• –Keeps revisions grounded in source trajectories to reduce feedback leakage
• –Adds a revision-agent call for each retrieved reflection, increasing latency and token consumption