When to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents
We introduce RECAP, a framework that combines matched context interventions, execution-relational prediction, and independent policy certification.
ProofPaper ↗
Key points
- Long-horizon agents learn from failure, but the value of a failure record can change as execution progresses.
- We study when retaining, compressing, or removing the same history block improves subsequent execution relative to full-history KEEP.
- Matched environment states and rollout seeds measure success and continuation prompt-cost differences; prefix-only relations between a failure block and the current execution propose promising edits.
- The framework provides an execution-level reference for deciding when to forget: retain history by its value to what comes next.
Sources (1)
- [1]When to Forget: Risk-Controlled Context Adaptation for Long-Horizon AgentsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 04:01 AM
We introduce RECAP, a framework that combines matched context interventions, execution-relational prediction, and independent policy certification.
Long-horizon agents learn from failure, but the value of a failure record can change as execution progresses.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026DLite: History-free Speculative Decoding for Agentic AI
- Oct 16, 2026Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design
- Oct 16, 2026HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
- Oct 15, 2026Anthropic gives outside AI watchdog “employee-like” access
- Oct 15, 2026Velox: Spatial Memory Makes World Models Efficient
- Oct 14, 2026mistralai/Timestral-209M