filesystem-context

Offload large tool outputs to the filesystem to prevent agent token bloating by using a durable storage layer for intermediate results. This approach keeps context windows lean while preserving data for on-demand retrieval.

Persist agent plans across context window refresh cycles by implementing plan persistence within a filesystem scratch pad. This ensures your multi-turn workflows survive restarts without losing intermediate task progress.

Share memory across multiple agents without heavy message passing by utilizing a shared filesystem context layer. Cross-agent information sharing is achieved by writing and retrieving scratch data durably from this common storage.

On-demand loading of scripts and references works by storing these assets in the filesystem and retrieving them selectively only when needed. This mechanism reduces baseline context usage and supports scalable agent workflows.

You need a scratch pad manager for multi-turn agent workflows when your tasks require persistent intermediate storage and cross-session memory. It manages dynamic context offloading to keep context windows from overflowing.

This approach to context management requires no external dependencies, operating entirely through filesystem scripts and references. It leverages your existing durable storage layer to handle tool outputs and plan persistence.