Beads Gives AI Coding Agents Persistent Memory With 26K GitHub Stars
Beads gives coding agents a structured, versioned task graph so they stop forgetting what they were doing mid-project
Takeaways
- What it is: Beads is a free, MIT-licensed CLI tool that gives coding agents a persistent, dependency-aware task graph backed by a versioned SQL database.
- The problem it solves: Agents lose task state when context windows fill up; markdown plans are unstructured, fragile, and cause Git conflicts in multi-agent setups.
- Core tech: Powered by Dolt, a Git-like SQL database with cell-level merge, native branching, and push/pull sync across machines and agents.
- Key features: Dependency graph, hash-based collision-free IDs, semantic memory decay (compaction), atomic task claiming, hierarchical epics, and stealth mode for shared repos.
- Multi-agent ready: Server mode supports concurrent writes; atomic
bd update --claimprevents two agents from picking up the same task simultaneously. - Install:
brew install beadsor via install script; supports Claude Code, Codex, Cursor, and more viabd setup <agent>.
Every developer who has handed a long coding task to an agent knows the feeling: you come back after a few hours, the context window has rolled over, and the agent has either forgotten half the plan or started redoing work it already finished. Beads is an open-source CLI tool built to fix exactly that. With over 26,000 GitHub stars, it has clearly struck a nerve.
Why Markdown Plans Fall Apart
AI coding agents hit two concrete walls on long tasks. Context windows truncate or forget earlier steps, and plans written in markdown have no structure an agent can reliably query or update. The instinct is to dump everything into a PLAN.md file, but that breaks down fast for three reasons:
- Markdown is plain text. "Updating status" means editing a file, which often ends with the agent deleting half the plan or creating Git conflicts when two agents work in parallel.
- A flat list carries no dependency information. An agent may not grasp that Task C requires Tasks A and B to finish first, leading to wrong execution order or skipped work.
- Agents have no persistent memory between sessions. Developers end up manually re-supplying context every time a new session starts.
Beads replaces markdown plans with a dependency-aware task graph backed by a version-controlled SQL database, giving agents the persistent state they need to handle long-horizon work across multiple sessions, branches, and collaborating agents.
A Database That Thinks Like Git
The storage layer is Dolt, a SQL database with Git semantics built in. Every write is a commit you can branch, merge, and roll back. Dolt's cell-level merge lets two agents update different fields of the same task row simultaneously without conflict, which is the property that makes concurrent multi-agent work tractable.