SDLC Agent-Skill Wiki

bd remember

2 min read

bd remember "insight" · bd recall <key> · bd memories · bd forget <key> — “store a memory that persists across sessions and account rotations”, auto-keyed from the content or pinned with --key, and injected into every session by beads-prime without anyone loading it.

bd remember "always run tests with -race flag"
bd remember "Dolt phantom DBs hide in three places" --key dolt-phantoms
bd remember "auth module uses JWT not sessions" --key auth-jwt
bd recall dolt-phantoms

The README’s instruction to agents is the sharp part: “Use bd remember for persistent project memory; do not create MEMORY.md files.” That is the same argument beads-todo makes about markdown checklists, aimed at knowledge instead of tasks — a file an agent writes is a file that rots, is not queryable, is not merged, and is not guaranteed to be read next time. A memory in the store is all four.

This is why the wiki’s node for this layer is store and not tracker. Beads is an issue tracker that also ships a durable knowledge base, keyed, mergeable, and auto-primed; strip the issues away and what is left is warren’s .mulch/ with a different CLI. The layer holds state, and work items are only one kind.

Against the process-layer implementations of pattern-knowledge-compounding the difference is who is responsible for reading it back. ce-compound and gstack-learn write lessons to a corpus and trust a later session to consult it; seeds-plan-prompt pulls the relevant subset into a planning prompt on demand. Beads pushes the whole set into context at session start unconditionally — cheapest to get right, and it scales badly on purpose, which is what beads-compact is for.

Maps to no canonical SDLC stage: storing an insight performs no lifecycle step, and unlike ce-compound there is no harvesting pass — an agent calls remember when it notices something, so the deciding is in whatever skill told it to.

See Also

Graph View

Backlinks