agentfs: Store and Audit AI Agent State in SQLite

Overview

AgentFS is a storage layer for AI agents. It puts an agent’s files, key-value state, and tool-call history in a SQLite database, so the agent’s state can be queried, copied, and moved as a single file.

It is aimed at developers building agents who need more than a temporary working directory: the database gives them an auditable record of changes and a way to snapshot or restore state. SDKs provide programmatic access, while the CLI can inspect and mount filesystems. AgentFS is a storage and state-management tool, not a complete agent framework or a substitute for a separate security boundary such as a container.

Key Features

• Store files and directories in a POSIX-like filesystem backed by SQLite.
• Keep agent key-value state alongside files in the same database.
• Record tool calls with inputs, outputs, status, and timing for later inspection.
• Query agent history and filesystem data using SQL.
• Snapshot or copy the database to preserve or restore an agent’s state.
• Access the storage through TypeScript, Python, and Rust SDKs.
• Inspect files and timelines with the CLI, or mount a filesystem through FUSE on Linux and NFS on macOS.

Use Cases

• Agent developers can retain tool-call history and state to investigate why a run behaved a certain way.
• Teams testing agents can copy a database to preserve a run’s state and reproduce experiments from that snapshot.
• Developers running multiple coding agents can use filesystem-level isolation for each agent’s files, including untracked files.
• Applications in environments without block-device mounting can use an SDK to manage agent state through the database interface.

Getting Started

Install the CLI and initialize a filesystem:

curl -fsSL https://agentfs.ai/install | bash
agentfs init my-agent

See the README and User Manual for SDK setup and further usage.

License in Plain Words

No standard license was detected. Without a license, the default copyright rules apply: you may read the code, but not reuse it. Check the repository or ask the authors.

Alternatives

• cortex-mem: Cortex-mem focuses on long-term memory extraction and semantic search, rather than storing an agent's filesystem and tool-call history.
• memvid: Memvid packages content, indexes, and metadata for local retrieval in an .mv2 file, rather than providing a SQLite-backed agent filesystem.
• engram: Engram retains project knowledge in SQLite and exposes it through MCP, a CLI, and APIs, rather than storing files and tool-call history.

Considerations

The README labels AgentFS as beta and cautions against relying on it with production data without backups. Health data reports no commits in the last 90 days and no issues closed or pull requests merged in that period, while releases continued during the past year. The project has 22 contributors and is not flagged as contributor-concentrated. SDKs are available in three languages, but the requirements data does not specify runtime prerequisites. The README states that the project is MIT licensed.