ams-memory-search
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-11-07
- Stars
- ★ 0
- Forks
- 1
03 SKILL.md 正文
Memory Search
Overview
Agents without memory start every conversation from zero. Memory systems store durable facts, decisions, preferences, and history — searchable when needed.
Effective memory search: store well (atomic claims, sourced, dated), retrieve with multiple query phrasings, and verify before acting on old memories.
The pattern: write memories as you learn, search before deciding, and update when facts change.
When to use
- Agent needs prior context about the user or project
- Recalling past decisions and their reasoning
- Finding what was learned in earlier sessions
- Verifying a fact before acting on it
- Maintaining continuity across long-running work
Core concepts
- Atomic claims. Store one fact per memory entry: who said it, when, the exact claim. Atomic entries retrieve precisely; paragraphs retrieve vaguely.
- Source and date. Every memory notes its origin (conversation date, document) and timestamp. Old memories need age-checking before use.
- Multiple query phrasings. Search with 2-3 different wordings of the same question. Retrieval is fuzzy — varied phrasings catch what one misses.
- Verify before acting. Retrieved memories are leads, not gospel. Check the source context for anything consequential before acting on it.
- Update, don't just add. When facts change, update or supersede the old memory. Contradictory memories without resolution poison future retrieval.
- Salience and decay. Not all memories matter equally. Boost frequently-used, correct memories; let stale, unused ones fade. Memory needs gardening.
- Scoped retrieval. Search the right store: user preferences vs. project facts vs. session history. Scoped search beats global noise.
- Provenance chains. For important claims, keep the chain: memory -> source conversation -> original evidence. Auditability matters for trust.
Practical workflow
- Capture as you learn. Durable facts, decisions, preferences -> memory immediately, in atomic form with source and date. Don't batch 'later.'
- Search before deciding. Before acting on prior work, preferences, or history: search memory with 2-3 phrasings. Make it a habit, not an afterthought.
- Read the context. For top hits, pull the surrounding source context. A claim without context is easy to misapply.
- Act on verified memories. Use verified memories to inform decisions; note the memory as the basis so the reasoning is traceable.
- Update on change. New information contradicting a memory? Update the entry (or mark superseded) immediately. Stale memories are worse than none.
- Garden periodically. Review memories: still true? Still relevant? Merge duplicates, retire the dead. Monthly for active stores.
- Handle conflicts. Two memories disagree: check sources and dates, prefer newer + better-sourced, record the resolution.
- Respect privacy. Memories contain personal data. Scope access, never leak across users, and honor deletion requests completely.
Common pitfalls
- Never searching. Having memory but deciding from scratch each time. The search habit is the whole value proposition.
- Single phrasing. One query, no hits, giving up. Retrieval needs 2-3 varied phrasings — different words catch different entries.
- Acting on stale memories. Using a year-old preference as current fact. Check dates; verify anything consequential.
- Memory hoarding. Storing everything verbatim. Uncurated memory becomes unsearchable noise — atomic, salient, gardened.
- Contradictions unresolved. Old and new memories disagreeing silently. Resolve explicitly or retrieval becomes a coin flip.
- No provenance. Facts without sources. When challenged ('where did you learn that?'), sourceless memories collapse.
- Cross-user leakage. One user's memories surfacing for another. Scope strictly; this is a privacy failure, not a feature.
- Treating memory as truth. Memory is a lead, not evidence. Verify against sources for anything that matters.
04 同仓库更多 Skills
web-scrape
Turn web pages into LLM-ready markdown via the Firecrawl API (renders JS, anti-bot handling). Use when browser_open/fetch fails on JS-heavy pages, or when you need clean markdown/structured extraction from a site. Requires FIRECRAWL_API_KEY.
design-taste-frontend
tasteskill v2, the anti-slop frontend skill. Art direction for websites, distilled from a study of 247 curated gallery sites (One Page Love, Awwwards Sites of the Day). Deals researched directions (idea, composition, type, palette, photographs, section blueprints) from a library, enforces a strict no-slop house style with a real-browser inspector, and runs a look-and-fix loop on screenshots. Use this whenever you build, design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack (HTML/CSS, React, Next.js, Vue, Svelte, Astro, Tailwind) — also when the user says a page looks generic, templated, "AI-made", sloppy, or asks to make it look designed, premium, award-level or less like every other site. Not for dashboards or app screens.
test-driven-development
Use when implementing any feature or bugfix, before writing implementation code
systematic-debugging
Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes
requesting-code-review
Use when completing tasks, implementing major features, or before merging to verify work meets requirements
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).