Introducing Beads: A coding agent memory system

The launch announcement. Steve had been vibe coding "like a madman for forty days and forty nights" — on the beach in Cabo San Lucas with Thorsten Ball, at 60mph on the freeway to Bellingham ("stupid and dangerous as hell. Didn't care"), at the mall with his wife Linh hiding behind the laptop. The original target was a grand orchestration engine named vibecoder; two foundational design mistakes (Temporal-native architecture, plus a Master Plan markdown-file hierarchy that ballooned to 605 obsolete plans) totaled the codebase by week five, and he burned it down. What survived was a git-backed issue tracker for coding agents. The diagnosis: 95 to 99% of interactions with coding agents could be handled by a properly briefed model — the bottleneck isn't planning or tool use, it's memory. Agents forget between sessions, rebuild the world from markdown each restart, and are terrible at curating those markdowns. Beads sits between the agent and the long-horizon work, holding issues, blockers, dependencies, and discoveries in a real schema (parent/child, blocking, discovered- from); one line in AGENTS.md or CLAUDE.md points the agent at bd and "you'll never lose track of work items again." Backed by git as JSONL, naturally distributed, with AI-mediated merge resolution.

This is the first of the four Beads launch essays — Beads hit a thousand GitHub stars in six days, with The Beads Revolution, Beads for Blobfish, and Beads Blows Up following.