Long-horizon agents record failures during execution, yet the value of a given failure block changes as time progresses. We investigate when retaining, compressing, or discarding the same history block can improve subsequent execution compared to keeping the full history (KEEP). To this end we introduce the RECAP framework, which consists of matched context interventions, execution‑relational prediction, and independent policy certification. Matched environment states and rollout seeds measure differences in success probability and continuation prompt cost; prefix‑only relations between a failure block and the current execution propose promising edit candidates. Independent task bundles then select a policy under a fixed measurement protocol, subject to harmful‑intervention and non‑positive‑benefit risk constraints. In a five‑seed longitudinal study we keep the failure block and its summary fixed: compression incurs extra cost at early states but becomes a cost‑saving intervention at later states. Across five backbone models RECAP yields positive paired‑success savings and exhibits backbone‑specific reliability‑efficiency trade‑offs. On Hunyuan‑A13B it edits 43.24% of retained checkpoints while matching KEEP’s 91.89% continuation success and reducing unconditional mean continuation prompt cost by 4,619.8 tokens (2.13%). The framework offers an execution‑level reference for forgetting decisions: retain history according to its value for what follows.
Review