Scroll Manages Long-Horizon AI Agent Context with Executable Environments.

Large language model agents often struggle with long-running tasks because their context windows are limited. Traditional methods either compress past interactions or extract specific information, which can lead to losing crucial details for future needs. Scroll addresses this by creating an "executable Session Environment" for each agent session. This environment uses an append-only Event Log and a persistent Python kernel. Instead of repeatedly embedding all information into the prompt, Scroll binds tool outputs, retrieved history, and derived states to variables within the kernel's namespace. The LLM can then use code to search, materialize, and transform this session state, with only explicitly printed projections entering its immediate working view. When the working view approaches its limit, older information is evicted but remains recoverable. An eviction index helps the agent navigate directly to these evicted regions in the Event Log, rather than searching the entire history. This programmatic approach to context management leverages the LLM's improving coding abilities and has shown significant performance gains on long-horizon memory benchmarks.