harden-memories

Harden Memories

Audit the project's bd remember entries and recommend which to prune down to the few that genuinely earn their cost. bd prime injects every bd remember entry in full at session start (they are persisted in the beads Dolt database), so entries accumulate as a silent, scaling token tax — conventions drift, skills ship, and facts graduate to CLAUDE.md / the auto-memory MEMORY.md / Basic Memory, but nothing automates the cleanup until a manual pass. Periodic audits (≈ every 4 sprints) reliably recover ~40% of that token budget.

This skill is READ-ONLY by design. It audits and proposes — it never edits a durable file, never writes to Basic Memory, and never runs bd forget itself. For each entry it presents a recommendation plus the exact command(s) for you to review and run. This deliberately keeps the irreversible step (bd forget, which destroys the only copy of a fact) and the migration write under your control — no automated "verify-then-delete" can be trusted to gate an irreversible delete (a Basic Memory write can fail schema_validate asynchronously; an Edit can land in the wrong file; an append can hit the wrong section). Proposing keeps you in the loop for both halves.

Scope — strictly the bd remember store. NOT Anthropic's Auto Dream / /dream (which prunes Claude Code auto-memory files, a different store), and NOT Basic Memory graph hygiene (defer that to vp-knowledge's graph-maintenance skills). It audits bd remember and nothing else.

Beads availability

Beads is available iff a .beads/ directory exists and command -v bd succeeds; this component is Tier B per CLAUDE.md ### Beads-availability convention. The bd remember store only exists when beads does — with no beads there is nothing to audit. Stop cleanly, naming the missing predicate (no .beads/ directory, or bd not on PATH). There is no beadless equivalent to redirect to.

Workflow

1. Read the current memories

bd memories --json