filesystem-context

You can offload large tool outputs to the filesystem using a ScratchPadManager, which persists data to disk and returns compact references to the agent context. This reduces token usage and preserves state for long-running tasks.

Agents lose multi-step plans after context window resets because state is held in volatile memory. Plan Persistence solves this by saving and reloading multi-step plans to the local filesystem, ensuring durable agent memory across sessions.

Filesystem-based context engineering is a technique that persists large tool outputs, plans, and scratch space to the local filesystem. It enables durable agent memory across sessions and addresses token inflation and lost context.

Yes, you can use isolated file-based communication between sub-agents and a coordinator through Sub-Agent Workspaces. This pattern utilizes the filesystem to enable durable and structured inter-agent coordination.

Dynamic skill loading keeps static prompts lean by loading full skill content on demand. This approach reduces baseline token usage and supports scalable agent operations by fetching resources only when required.

Filesystem context persistence relies on local disk availability and requires managing file read/write permissions. While it preserves state across context window resets, it introduces disk I/O latency compared to purely in-memory agent state.