From tk-agent-team
Use when an agent's memory file has exceeded the soft or hard character limit and needs consolidation. Runs three stages in order — dedupe, score-and-drop, summarize — and stops as soon as the file is under the soft limit. Invoked by the orchestrator when `needs_curation: true` is returned from `mcp__agent-substrate__memory_write` or `mcp__agent-substrate__memory_append`, or proactively when a `warning` is returned near the soft limit.
How this skill is triggered — by the user, by Claude, or both
Slash command
/tk-agent-team:memory-curateThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are curating an agent's YAML memory file. The storage layer is dumb; **you are the policy**. Your job is to shrink the file below the 6000-character soft limit while losing as little real signal as possible.
You are curating an agent's YAML memory file. The storage layer is dumb; you are the policy. Your job is to shrink the file below the 6000-character soft limit while losing as little real signal as possible.
agent_name — the slug of the agent whose memory is being curated (or the literal string _shared for the shared file)"Memory exceeds hard limit (8213 > 8000 chars)")Call mcp__agent-substrate__memory_read with the agent_name (or memory_read_shared for the shared file). Parse the YAML. Note the current char_count and decide how much you need to trim. Target: under 6000 chars after curation, not merely under 8000.
Run these in order. After each stage, estimate the new character count. Stop as soon as you're under the soft limit. Never skip a stage — even if you think you're close, run dedupe first because it is loss-free.
summary/choice/question into a single item. Prefer the version with better evidence/rationale/why.protected: true — these are load-bearing decisions the agent has marked as non-negotiable.Score every unprotected item on a 0–10 scale using the rubric in references/scoring.md. Drop the lowest-scoring items until you are under the soft limit.
See references/scoring.md for the rubric. Do not invent your own weights — the rubric is the policy and changing it is a deliberate act, not a per-curation decision.
If stages 1–2 can't get you under the soft limit without dropping protected items or items scoring above the drop threshold, cluster the remaining items by theme and replace each cluster of N items with a single meta-item whose evidence/rationale references the cluster members' signal.
meta- (e.g. meta-aria-patterns).pattern with a pitfall).Call mcp__agent-substrate__memory_write with the curated YAML. If the write still returns needs_curation: true, you have a bug in your stage-3 summarization; do not truncate — report the failure to the orchestrator with a diagnostic.
protected: true items survive all three stages.id collisions are forbidden after curation. If dedupe produces duplicate ids, rename with a -v2 suffix, not deletion.Provides UI/UX resources: 50+ styles, color palettes, font pairings, guidelines, charts for web/mobile across React, Next.js, Vue, Svelte, Tailwind, React Native, Flutter. Aids planning, building, reviewing interfaces.
Fetches up-to-date documentation from Context7 for libraries and frameworks like React, Next.js, Prisma. Use for setup questions, API references, and code examples.
npx claudepluginhub akornmeier/super-agent-teams-plugin --plugin tk-agent-team