What problem does it solve?
Deep agents need both short-term thread-scoped state and long-term persistent memory plus controlled filesystem access; this Skill documents how to configure and use pluggable backends to provide those capabilities safely. It clarifies when files are ephemeral, when they persist across sessions, and how to route different paths to different storage systems to avoid accidental data loss or leakage.
Core Features & Use Cases
- StateBackend (short-term): ephemeral, thread-scoped storage for temporary working files and drafts.
- StoreBackend (long-term): persistent storage that survives threads and sessions and requires a store instance (e.g., InMemoryStore for testing, PostgresStore for production).
- CompositeBackend (hybrid routing): route path prefixes to different backends with longest-prefix matching to mix ephemeral and persistent storage.
- FilesystemBackend & Middleware: controlled disk access for local development with tools like ls, read_file, write_file, edit_file, glob, and grep, and a virtual_mode option to restrict path escapes.
- Use Cases: save user preferences across sessions, keep transient editor drafts per thread, enable human-in-the-loop file edits during local testing, and migrate production storage to a durable Postgres store.
Quick Start
Configure an agent with a CompositeBackend that routes /memories/ to a persistent store and uses a StateBackend for transient files so that drafts remain ephemeral while user preferences persist.
