Splitting agent memory in two lifts tau-bench 15.6 points

Recuris, a memory architecture described in a paper posted to arXiv (2608.24876), splits an AI agent's memory into two separate stores. One tracks progress on the current task. The other holds skills the agent has learned over time. Most agent memory systems mix the two, so skill selection gets diluted by everything else in the agent's history.

On Opus 5, the split adds 15.6 points on tau-bench, a benchmark built around multi-step agent tasks. It adds 17.8 points on GPT-5.6 Sol, and on the longest-horizon tasks tested it reaches 32.2 points. Across 37 combinations of models and benchmarks, the approach won on 35.

The paper also adds a meta-agent that watches for failures and traces each one to a specific memory component, Working or Experiential. With the blame pinned to one store, common failure modes drop as much as 80%, instead of forcing a retrain of the whole system.

None of this needs a new model. It's a change to how memory gets organized, and it's cheap enough for any team building agent memory to test.