Beads - Persistent Task Memory for AI Agents

Graph-based issue tracker that survives conversation compaction.

Overview

bd (beads) replaces markdown task lists with a dependency-aware graph stored in git.

Key Distinction:

• bd: Multi-session work, dependencies, survives compaction, git-backed

Decision Rule: If resuming in 2 weeks would be hard without bd, use bd.

Operating Rules

• Treat live bd reads as authoritative. Use bd show, bd ready, bd list, and bd export to inspect current tracker state. Do not treat .beads/issues.jsonl as the primary decision source when live bd data is available.
• Treat .beads/issues.jsonl as a git-friendly export artifact. If the repo tracks .beads/issues.jsonl and you mutate tracker state, refresh it explicitly with bd export -o .beads/issues.jsonl.
• After closing or materially updating a child issue, reconcile the open parent in the same session. Update stale "remaining gap" notes immediately, and close the parent when the child resolved the parent's last real gap.
• If bd ready returns a broad umbrella issue, do not implement directly against vague parent wording. First narrow the remaining gap into an execution-ready child issue, then land the child and reconcile the parent.
• Normalize stale queue items instead of silently skipping them. Rewrite broad or partially absorbed beads to the actual remaining gap.
• Use this post-mutation sequence when tracker state changed:

bd ...                              # mutate tracker state
bd export -o .beads/issues.jsonl    # if tracked in git
bd vc status
bd dolt commit -m "..."             # if tracker changes are pending
bd dolt push                        # only if a Dolt remote is configured

Prerequisites

• bd CLI: Version 0.34.0+ installed and in PATH
• Git Repository: Current directory must be a git repo
• Initialization: bd init run once (humans do this, not agents)

Examples

Skill Loading from $vibe

User says: $vibe

What happens:

1. Agent loads beads skill automatically via dependency
2. Agent calls bd show <id> to read issue metadata
3. Agent links validation findings to the issue being checked
4. Output references issue ID in validation report

Result: Validation report includes issue context, no manual bd lookups needed.

Skill Loading from $implement

User says: $implement ag-xyz-123

What happens:

1. Agent loads beads skill to understand issue structure
2. Agent calls bd show ag-xyz-123 to read issue body
3. Agent checks dependencies with bd output
4. Agent closes issue with bd close ag-xyz-123 after completion

Result: Issue lifecycle managed automatically during implementation.

Troubleshooting

Reference Documents

• references/ANTI_PATTERNS.md
• references/BOUNDARIES.md
• references/CLI_REFERENCE.md
• references/DEPENDENCIES.md
• references/INTEGRATION_PATTERNS.md
• references/ISSUE_CREATION.md
• references/MOLECULES.md
• references/PATTERNS.md
• references/RESUMABILITY.md
• references/ROUTING.md
• references/STATIC_DATA.md
• references/TROUBLESHOOTING.md
• references/WORKFLOWS.md

Local Resources

references/

• references/ANTI_PATTERNS.md
• references/BOUNDARIES.md
• references/CLI_REFERENCE.md
• references/DEPENDENCIES.md
• references/INTEGRATION_PATTERNS.md
• references/ISSUE_CREATION.md
• references/MOLECULES.md
• references/PATTERNS.md
• references/RESUMABILITY.md
• references/ROUTING.md
• references/STATIC_DATA.md
• references/TROUBLESHOOTING.md
• references/WORKFLOWS.md

scripts/

• scripts/validate.sh