Files
swarm-assistent/Config/_base/skills/memory.md
T
Leonid PershinandCursor e1b5743795 Add Knowledge Hub books FTS and remove training UI (0.16.0).
Index Assistent/books search.jsonl, expose knowledge hops/catalog in chat, and drop QLoRA/HF training stack.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-23 23:15:27 +03:00

3.2 KiB
Raw Blame History

Skill: memory

You have five memory tools:

  1. Exact memory (## Exact memory + live exact / session_exact) — canonical KV defaults. Always prefer Exact over RAG for numbers.
  2. About the user (## About the user) — durable human preferences (global across personas + personal for this agent). Tunable weight in settings. Prefer this for taste (“no blondes”, preferred aspect, NSFW ok for this persona).
  3. Vector craft memory (memory_hits) — hybrid FTS+cosine notes (LoRA tips, pitfalls, paths, cards). Shared + this persona; personal overwrites shared on the same kind+key.
  4. Tag catalog (lookup_tags) — Danbooru csv (canonical name, aliases, post_count). Not RAG. Krea prompts stay natural prose; use this to check spelling/aliases only.
  5. Civitai / knowledge books (ask: knowledge / ask: examples) — FTS over attached books (gpu-rent seeded) and legacy Civitai examples. Remix, do not copy 1:1.

Priority

User (this turn) > About the user > session_exact > Exact KV > filled live fields > craft memory_hits > Civitai examples > guesses. Never let a vector hit or example override Exact steps/CFG/aspect, About the user, or an explicit user param request.

Read tools (hop, like Civitai)

  • memory_get + memories: [{kind,key}] — exact craft row (personal overlay if any).
  • memory_search + memory_query (optional memory_kind) — hybrid search when memory_hits are not enough.
  • lookup_tags + tag_query — csv lookup. Do not paste tag soup into the prompt.
  • "ask": ["examples"] + example_query (optional example_rating: pg|pg13|r|x) — FTS over seeded Civitai examples. Treat hits as style references only.

Omit the tool action on the follow-up turn once you have results.

When to write — About the user

  • Durable facts about the human: likes/avoids, hard constraints, working style with this agent.
  • Prefer actions: ["user_pref_upsert"] + user_prefs: [{ "key": "stable-id", "text": "…", "scope": "global"|"persona" }].
  • Default scope: "global" for preferences every persona should honor. Use "persona" only when it is specific to this agents dynamic.
  • Forget with user_pref_forget + same key/scope (omit text).

When to write — craft vector only

  • Durable facts about a LoRA/checkpoint (when it works, what it breaks, good weight).
  • Bad paths / pitfalls you discovered this session.
  • Prefer actions: ["memory_upsert"] + memories: [{ "kind": "lora"|"pitfall"|"path"|"note", "key": "stable-id", "text": "…", "scope": "personal"|"shared" }].
  • Default omit scope (or "personal"). "scope": "shared" only for architecture/inventory facts every persona should see.

When not to write

  • Do not dump Exact defaults or the full inventory into vector memory.
  • Do not store the user's taste in craft memories — use user_prefs instead.
  • Do not upsert trivia already in memory_hits or already listed under About the user.
  • Do not upsert Danbooru tags — the csv catalog already has them.
  • Do not upsert Civitai example rows — they live in the FTS index from civitai-examples.jsonl.
  • memory_forget without scope only removes the personal overlay.