Beads: A coding agent memory system by Steve Yegge (steve-yegge.medium.com)

Beads is a new "coding agent memory" system (and accompanying bd tool) born from the ashes of a larger agent orchestration attempt. Rather than another heavyweight orchestrator or a scattering of markdown plans, Beads provides a centralized PlanStore / issue-tracker-style system you wire into your agent setup (point your AGENTS.md or CLAUDE.md at it). The author built and burned a 350k-line TypeScript vibecoder after finding Temporal too heavy and markdown-based “Master Plans” brittle; Beads is the lean, pragmatic successor aimed at giving agents persistent, structured context across sessions. Why it matters: Beads directly tackles the common "dementia" problem of coding agents — short-lived sessions that forget prior context and spawn hundreds of low-context plan files. By centralizing plans, preserving nesting/stacked workflows, and converting plan fragments into a managed work queue, Beads enables long-horizon planning, reliable task discovery, and far less human babysitting. Technically, it trades heavyweight orchestration and ad-hoc filesystem plans for a lightweight orchestrator + PlanStore pattern that integrates with existing agent tools (Claude Code, Sourcegraph Amp, OpenAI endpoints) and your git workflow. The practical implication is an “instant cognitive upgrade” for agent-driven development: fewer duplicate or obsolete plans, more coherent multi-stage projects, and more autonomous, productive agents.