filesystem-context

You can persist agent context to the filesystem by offloading large tool outputs to scratch files and storing long-horizon plans on disk. This extends agent memory beyond token limits by returning compact references instead of keeping full context in memory.

Filesystem context engineering is the practice of offloading large context pieces to disk for later reuse. It reduces token pressure and enables persistent agent memory by using a stable filesystem layout for scratch pads, plan persistence, and per-agent workspaces.

Set up a ScratchPadManager to write large tool outputs to scratch files on disk, returning compact file references for the agent context. This process offloads bulky data from the context window while keeping the information accessible for later reference.

Yes, you can create per-agent file-based workspaces to isolate outputs and enable coordination in multi-agent collaboration. These isolated filesystem layouts allow multiple agents to operate independently while maintaining structured access to their respective persistent data.

The best way to persist long-horizon agent plans is writing them to disk using structured plan storage conventions. This ensures tasks survive context window refreshes by allowing the agent to reload the persisted plan from the filesystem when needed.

Dynamic skill loading retrieves full skill content on-demand from indexed SKILL.md files when needed. This approach minimizes token usage by keeping skill metadata lightweight until the complete skill instructions are explicitly required by the agent.