agentic-context-management

This is the official repository for the paper "ACM: Agentic Context Management for Long-Horizon Tasks" [https://arxiv.org/abs/2607.23809]

We propose a post-training pipeline that teaches a model to manage its own context on long-horizon agentic tasks. It has two key properties. First, it is agent-native: the agent decides when to compress context itself, instead of being forced by an external trigger. Second, it is lossless: discarded context is written to disk and kept permanently, so the agent can go back and retrieve any earlier information when it needs it.