Beads - A memory upgrade for your coding agent
bd) is a Dolt-powered distributed graph issue tracker that gives coding agents persistent, structured, version-controlled memory across sessions.| Rating | Summary | |
|---|---|---|
| Quality | solid (19/24) | Actively released, extremely well-adopted (24k stars in 7.5 months), thorough documentation, and clean CLI design — held back only by 215 open issues and absent test/CI evidence in the README. |
| PAI Relevance | integrate (0.88) | Fills PAI's gap in versioned, dependency-aware agent task tracking; the bd CLI subprocess interface, JSON output, and offline-first design fit PAI's stack directly. |
19/24 — maintained / well-documented / solid
Failed:
.github/workflows/.Passed:
archived: false.Failed:
Passed:
curl | bash install and bd init walkthrough.https://gastownhall.github.io/beads/ and multiple docs/*.md pages.Failed:
Passed:
go.mod is mandatory for any published Go module and is implicitly present.| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 2 | Beads introduces three patterns directly applicable to PAI's memory subsystem: (1) Dolt-backed versioned task graphs with cell-level merge for multi-agent conflict resolution, (2) semantic memory decay/compaction to manage context window budgets (directly applicable to PAI's WORK/ and LEARNING/ memory management), and (3) bd remember + bd prime as a reinjection loop for persistent project insight — a concrete implementation of a pattern PAI lacks. |
| Integration Readiness | 2 | bd is a language-agnostic CLI binary (Go) installable via brew install beads or install script; it emits structured JSON output and works entirely offline. PAI can subprocess bd commands from any skill or hook with no adapter layer. An MCP server (beads-mcp) is also available for direct agent tool-call integration. |
| Overlap Risk | 1 | Partial overlap with PAI's ISA skill, which manages ideal-state artifact tracking and task decomposition. ISA does not support dependency graphs, multi-agent conflict resolution, versioned history, or memory compaction — so overlap is meaningful but not deep. |
| Gap Fill | 2 | PAI has no versioned, graph-structured task tracker with dependency resolution, atomic claiming, or multi-agent branch isolation. File-based memory in WORK/ and LEARNING/ covers persistence but not structured task graphs or conflict-free concurrent writes — a clear functional gap Beads fills. |
Composite: 0.88
Capture-to-Knowledge Pipeline (design + validation phase): Run go install github.com/gastownhall/beads/cmd/bd@latest, then bd init in the pipeline repo and decompose the remaining stages (clean-room validation → production ingestion → downstream consumers) into dependency-linked Beads tasks using bd task add --after <id>. At the start of each Claude Code session, the agent calls bd task list --ready instead of re-reading design docs — every session resumes with a machine-queryable dependency graph rather than a re-explained markdown brief.
Aphoria Stage 3B+ and Conservancy editorial work: For any project that will require 3+ agent sessions to complete the next development stage, initialize a Beads database at the project root before handing off to Claude Code. Instruct the agent to use bd task add to decompose its work plan on first session and bd task done <id> as it progresses. When you return in a later session, bd task list --pending restores full task state in one command — replaces the fragile pattern of agents writing PLAN.md files that go stale or get overwritten mid-session.
Long-horizon agent work across any active project: For tasks that will span enough sessions to strain context budgets, configure Beads' semantic memory compaction threshold so completed tasks are summarized and pruned before older history crowds out current context. This is the only agent memory tool with a built-in context-budget governor — without it, agents in session 6 of 8 are quietly dropping early task rationale to stay within limits; with it, compaction is explicit and version-controlled via Dolt so you can audit what was pruned.
Category: Developer Tools & CLIs
In this category: casey--just (excellent, 21/24) — first entry in agent-specific task tracking within this category.
Standing: First repo in the vault addressing persistent agent task memory via a graph database; casey--just is the only other Developer Tools entry and solves a complementary (not competing) problem.
Density: 9/10 — Available: full README (8KB), repository metadata (stars, forks, issues, dates, topics, license, language), release history with version and date, landscape context and related repos. Missing: dependency manifest content (dep count unverifiable), CI configuration details, test file structure, full docs/ directory content.
github.com/steveyegge/beads rather than gastownhall/beads, suggesting the repository was renamed or transferred from a personal account — worth noting for provenance tracking.gastownhall--gastown (multi-agent workspace manager), suggesting a complementary two-tool ecosystem: gastown for agent orchestration, beads for agent memory.