memory-curator
Use at session-close when a session captured several new memories, or on-demand, to consolidate, deduplicate, prune, and structure the beads memory store. Triggers on "curate memories", "clean up memories", "memory sweep".
DollarDill · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add DollarDill/beads-superpowers --skill memory-curator技能说明
name: memory-curator description: Use at session-close when a session captured several new memories, or on-demand, to consolidate, deduplicate, prune, and structure the beads memory store. Triggers on "curate memories", "clean up memories", "memory sweep".
Memory Curator
Turn a session's raw bd remember notes into deduplicated, consolidated, well-structured
memories — and prune the pile — using bd over text already in context. No runtime, no embeddings.
Announce at start: "I'm using the memory-curator skill to consolidate and structure the memory store."
When to Use
- Session-close — when the session produced ~3+ new
bd remembercalls. Offered, never automatic. - On-demand — a full-store sweep:
Skill(beads-superpowers:memory-curator).
When NOT to Use
- Sessions with 0–2 new memories — not worth a pass.
- Mid-task — run at a clean stopping point, not while work is in flight.
Memory taxonomy
Two classes; procedural memory (how-to / workflow) lives in the skills, never the memory store.
- semantic — durable facts that stay true.
- episodic — time-bound records of what happened.
The @type is the routing decision. Classify once; the type deterministically sets store, injection, and lifecycle.
@type | store | injected at session start? | lifecycle |
|---|---|---|---|
semantic:lesson | memory | yes (salience≥4) | durable; consolidate near-dups |
semantic:root-cause | memory | yes (salience≥4) | durable; consolidate |
semantic:pattern | memory | yes (salience≥4) | durable; consolidate |
semantic:correction | memory | yes | durable (supersedes a wrong memory) |
semantic:research | deferred knowledge-bead | no | deferred bead; pointer to a research doc (metadata.doc); bd supersede on replacement |
semantic:design | deferred knowledge-bead | no | deferred bead; pointer to an ADR/spec (metadata.doc); bd supersede on replacement |
semantic:decision | deferred knowledge-bead | no | deferred bead; pointer to an ADR (metadata.doc); bd supersede on replacement |
episodic:continuation | memory | latest only | supersede on next |
episodic:done / cleanup / review | memory → retire | no | consolidate into a semantic fact, then drop; age-out (>30d) safety net |
Crisp routing definitions (the boundary that keeps determinism honest):
research/design/decision= a pointer whose detail lives in a doc/ADR you would re-open when relevant. Injecting it every session wastes context — route to a deferred knowledge-bead (§ Beads-native knowledge store).lesson/root-cause/pattern/correction= a standalone, actionable rule you want surfaced unprompted so you don't repeat a mistake (e.g. "bd worktree default path is ./, not .worktrees/"). Stays an injected memory.- Escape hatch: if a research/design item is genuinely a standalone reusable rule, classify it as a
lesson/pattern— you change the type, never the store directly.
Map a non-canonical prefix to the nearest canonical subtype — e.g. stress-test/plan-stress-test→design,
bug→root-cause, sdd→lesson, upstream→research, docs→pattern. If none fits, ask — don't
invent. If an extracted "memory" is really procedural, flag it for a skill — don't store it.
Memory header
Every memory keeps its existing key and carries one greppable header line:
@type=semantic:lesson @created=2026-06-28 @salience=4 @refs=<bead-id>,<memory-key> @tags=memory,curation
<self-contained fact body>
@type—<class>:<subtype>from the taxonomy — the subtype sets store/injection/lifecycle per the taxonomy table above.@created— ISO date.@salience— 1–5, best-effort.@refs— related bead IDs / memory keys.@tags— lexical filter.
The class makes the prune signal greppable (bd memories | grep '@type=episodic:');
@salience/@tags filter recall.
Beads-native knowledge store
Reference-class memories (research/design/decision) live as deferred knowledge-beads, not in memory. — a deferred bead is never auto-injected at session start, so pointers stay out of every session's context but keep persistence + Dolt sync.
-
Bead:
status=deferredwith a far-future--defer 2099-01-01— neverclosed(closed beads are GC-deleted at 90d).issue_typematches the subtype (research/design/decision); every knowledge-bead also carries the class-marker labelkbplus 1–3 topic labels from the controlled vocabulary (scripts/kb-label-vocab.txt). -
Body: the research doc / ADR stays on disk as the source of truth; the bead is the queryable index/pointer via
metadata.doc(display-only), with a distilled summary as the description:printf '%s' "<distilled summary: what this note establishes>" | \ bd create "<one-line summary>" -t <research|design|decision> -l kb,<topic-labels> \ --defer 2099-01-01 --metadata "$(jq -nc --arg d "<doc-path>" '{doc:$d}')" --body-file - --silent -
Retrieval:
bd list --label <topic> --status all(topic) andbd search "<kw>" --status all(keyword) — never metadata filters (broken inbd), neverfind-duplicates. -
Lifecycle:
bd supersede <old> --with <new>on replacement — the superseded bead closes and decays; the live pointer stays deferred. -
Move-out invariant (curator route step): write the deferred knowledge-bead → verify (
bd show <id>returns it) → thenbd forgetthe memory. Never forget first. Existence-check before writing (idempotent re-run). Run the secret/PII scan on the body first — flag for removal, never relocate a secret into a bead. -
Aging path: a cooled injected memory (low
@salience, or superseded) can retire into a deferred knowledge-bead too, not just a tombstone — same move-out invariant above (write → verify →bd forget), never a copy left behind in both stores.
The sweep
One pass. Input: the session (i
阅读完整来源 ↗