UNF - Filesystem Flight Recorder for AI Agents
UNF continuously records every file change on your filesystem, letting you rewind any file to any second. Open-source tool for developers working with AI coding agents.
TL;DR
TL;DR: UNF is a local, open-source filesystem watcher that records every file change in real time, letting you rewind any file to any second — useful when an AI agent wipes your workspace between context compaction windows.
What Is UNF?
UNF (stylised UNF*) is a filesystem flight recorder. It runs continuously in the background, watching your project directory for every file change, and stores a point-in-time snapshot you can restore from at any moment.
The problem it solves: when an AI coding agent rewrites files during a session, those changes live only in the agent’s context. Once the context compacts or the session resets, the agent forgets what it touched. If it then makes a conflicting edit, your original code is gone.
UNF sits below that workflow. It records the filesystem state at second-level granularity regardless of what tool made the change — Git, an editor, a script, or an AI agent.
The product page describes it directly:
Every file change, recorded in real time. Rewind any file to any second. No commits required.
UNF is open source under MIT / Apache-2.0.
How It Works
UNF has three moving parts:
01 WATCH — Uses FSEvents on macOS and inotify on Linux to monitor the filesystem. It respects .gitignore, skips binary files, and applies a 3-second smart debounce to avoid recording every keystroke. Resource usage is claimed at under 1% CPU and under 100MB RAM during active recording.
02 STORE — Snapshots are stored locally using BLAKE3 content hashing inside a SQLite database with ACID guarantees. Content-addressed storage means identical files are stored only once. All data stays local.
03 REWIND — Any file can be restored to its state at a specific point in time, by second. Manual pruning with unf prune reclaims disk space.
Practical Workflow
Start watching a project:
unf watch
Review a session after an agent has been active:
unf recap --since 3h
Inspect what a file looked like 30 minutes ago:
unf diff --at 30m path/to/file
Restore a file to its state 30 minutes ago:
unf restore --at 30m path/to/file
The unf diff output shows a standard unified diff, so you can read exactly what changed before deciding whether to restore.
Installation
macOS (Homebrew):
brew install --cask cyrusradfar/unf/unfudged
macOS + Linux (Homebrew CLI):
brew install cyrusradfar/unf/unf
Linux (Debian/Ubuntu):
curl -fsSLO https://downloads.unfudged.io/releases/latest/unf_amd64.deb
sudo dpkg -i unf_amd64.deb
From source (requires Rust 1.80+):
git clone https://github.com/cyrusradfar/homebrew-unf.git
cd homebrew-unf
cargo build --release
cp target/release/unf /usr/local/bin/
Limitations
UNF is a filesystem recorder, not a version control system. It does not track branches, manage commits, or sync across machines. If you need collaborative history, Git remains the right tool. UNF fills the gap for in-progress, uncommitted work that Git cannot capture.
The current version stores snapshots locally. There is no cloud sync or multi-device support.
FAQ
Q: Does this work on Windows? A: Currently no. UNF uses FSEvents on macOS and inotify on Linux. Windows is not supported at this time.
Q: How much disk space does it use?
A: It depends on your project size and change frequency. Use unf prune to manually clean up old snapshots and reclaim space.
Q: Does it interfere with editors or AI coding agents?
A: No. UNF only watches the filesystem — it does not instrument or interact with running processes. It respects .gitignore and skips binary files.
Q: Is the source code available?
A: The project is open source. The Homebrew tap at github.com/cyrusradfar/homebrew-unf hosts the installation tooling. License is MIT / Apache-2.0.
Conclusion
UNF solves a specific, painful problem that surfaces whenever developers work with AI coding agents in long sessions: loss of uncommitted state when the agent’s context resets. It is lightweight, local-only, and requires no infrastructure beyond installing a binary.
If you regularly work with AI agents that modify code across sessions — or if you have ever lost hours of work because “the agent said it had the right version” — UNF is worth a look.
Source and install instructions are at unfudged.io.
Source and Accuracy Notes
- Project page: unfudged.io
- License: MIT / Apache-2.0 (stated on project page)
- HN launch thread: news.ycombinator.com/item?id=44190017 — 137 points
- Source last checked: 2026-08-16
Related Posts
ai-setup
Recall – Persistent Memory for Claude Code via MCP Hooks
Recall gives Claude Code a permanent memory store that survives session restarts and context compaction. Four hooks capture and restore context automatically — with cloud SaaS or self-hosted options.
2/28/2026
dev-tools
Automotive Skills Suite for AI Engineering
Evaluate Automotive Skills Suite for APQP, ASPICE, HARA, safety-plan, and DIA workflows with setup notes, governance risks, and SME review guidance.
5/28/2026
dev-tools
awesome-agentic-ai-zh Roadmap Guide
Explore awesome-agentic-ai-zh as a Chinese agentic AI learning roadmap, with setup notes, track selection, study workflow, and evaluation guidance.
5/28/2026