Published
Context compaction for long-horizon agents
A method for compacting conversation history that preserves task-relevant context while bounding token growth across long sessions.
A. Researcher, B. Engineer
contextmemoryagents
Summary
We study how to keep long-running agent sessions within a fixed token budget without losing task-relevant context. The full writeup is in the PDF.
Results
Compaction holds task accuracy within 2% of the uncompacted baseline while cutting context tokens by 60% on sessions longer than 50 turns.