Beads: Memory for Your Coding Agents

In my previous post on Yegge, I covered his “divers and CNC machines” thesis: current agents burn context like oxygen, and the future is orchestrated swarms. His new project, Beads, is a step toward that future.

The problem it solves: coding agents have amnesia. Every session is 50 First Dates.

The Markdown Nightmare

Before Beads, Yegge spent a month drowning in markdown files. Hundreds of them, accumulating:

All they know is what’s on disk. If you got competing documents, obsolete documents, conflicting documents, ambiguous documents - they get dementia.

— Steve Yegge

Agents don’t know what’s current. They can’t distinguish between “we decided this yesterday” and “this was a brainstorm from three weeks ago.” Everything looks equally valid.

The solution: addressable work items. Every task gets an ID, a priority, dependencies, an audit trail. Not a wiki. Not scattered notes. An issue tracker.

Why Git

Beads stores everything in Git. This seems weird for an ephemeral issue tracker, but the reasoning is solid:

• Archived issues stay in history. Close it, archive it, delete it. Still recoverable.
• Works offline. No external service to depend on.
• Agents already know Git. No new integration needed.

There’s a SQLite cache for fast queries, but the source of truth is version-controlled markdown. The AI designed it that way when Yegge asked what it wanted.

Claude said ‘you’ve given me memory - I literally couldn’t remember anything before, now I can.’ And I’m like, okay, that sounds good.

— Steve Yegge

Current Work, Not Future Planning

Beads sits in a narrow space:

• Future work: vague, months out, lives in your backlog. Not Beads.
• Past work: done, documented, low risk. Not Beads.
• Current work: what you care about right now, what just finished and might break, what’s blocked. This is Beads.

The framework even prioritizes recently closed work higher by default. If you ship something and find P2 follow-ups, they’ll languish forever without a boost. Beads gives them one automatically.

Land the Plane

The pattern that sold me: “land the plane.”

At the end of every session, Yegge tells his agent: “Let’s land the plane.” This triggers a scripted cleanup:

1. Update Beads issues with progress
2. Sync the issue tracker
3. Clean up Git state (stashes, old branches)
4. Remove debugging artifacts
5. Generate a prompt for the next session

That last step is the key. The agent looks at Beads, finds the highest-priority unblocked work, and spits out a ready-to-paste prompt. Next session: copy, paste, go.

Their reward function biases them for checklists and acceptance criteria. Landing the plane, even if they’re low on context, they’re going to do a good job with it.

— Steve Yegge

The Numbers

Four weeks old. 29 contributors with merged PRs. “Tens of thousands” of users according to Yegge.

People have already built:

• Web UIs on top of Beads
• Monitoring systems for automated workflows
• Integrations with existing GitHub/Jira issues

The adoption is fast because the pain is real. Anyone who’s spent hours re-explaining context to their coding agent every morning gets it immediately.

What’s Next?

The holy grail: wake up, ask your agents “what’s next?”, and they know.

Not because you briefed them. Because Beads gives them memory. They can see the dependency graph, the priorities, what’s blocked, what just shipped. They orient themselves.

Man, that to me is the holy grail. And that’s what coding is going to be like next year.

— Steve Yegge

Running in Shoes

Yegge’s metaphor: using coding agents without something like Beads is running in socks. You get flexibility and some protection, but your feet hurt.

Beads is shoes. Opinionated, not appropriate for every situation, but damn useful when they fit.

The broader thesis is the same as my previous Yegge post: bigger context windows aren’t the answer. Better orchestration is. Beads is a piece of that puzzle: persistent memory that survives session boundaries.

Four weeks in, already essential to people’s workflows. Worth watching.