memory-search
/memory-search
Przeszukuje pamięć organizacji przed niebanalnym zadaniem: wcześniejsze decyzje, wyciągnięte lekcje, poprzednie prace i właścicieli.
- Autor
- ibrain-bvba ↗
- Repozytorium
- gutt-claude-code-plugin ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 14.09.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 2
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 14.09.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Przeszukuje pamięć organizacji przed niebanalnym zadaniem: wcześniejsze decyzje, wyciągnięte lekcje, poprzednie prace i właścicieli.
Oryginalny opis (EN)
Search organizational memory efficiently before any non-trivial task — an adaptive, relevance-gated first pass over the gutt knowledge graph, deepening only when needed. Use to recall prior decisions, lessons learned, past work, patterns, owners, or why something was done. Triggers on: previous, before, last time, decision, why did we, what do we know about, prior art, past work, lessons learned, have we done, did we already, history of, who worked on.
Jak działa
- Hard rules (non-negotiable — read first)
- When to use
- The search ladder
- Summary-first discipline
- Degradation
- References
Podgląd SKILL.md
---
name: memory-search
description: "Search organizational memory efficiently before any non-trivial task — an adaptive, relevance-gated first pass over the gutt knowledge graph, deepening only when needed. Use to recall prior decisions, lessons learned, past work, patterns, owners, or why something was done. Triggers on: previous, before, last time, decision, why did we, what do we know about, prior art, past work, lessons learned, have we done, did we already, history of, who worked on."
---
# Memory Search
How every agent should read the gutt knowledge graph: one strong first pass,
judged for relevance, deepened only when it falls short. This is the most-used
memory operation and the foundation skill — other memory skills
cross-reference it rather than restate it. Its job is to surface the answer as
fast as possible **when it exists**, and to say so plainly when it doesn't.
## Hard rules (non-negotiable — read first)
1. **One adaptive pass first.** Open with a single pass — `search_memory_nodes`
and `search_memory_facts` in parallel, on your best phrasing. Never open with
traversal, schema calls, or `fetch_lessons_learned`.
2. **Relevance gate.** Answer only from results that genuinely fit the question.
If nothing on-topic comes back, say **"no relevant memory found"** — never
assemble an answer out of loosely-matching distractors.
3. **Reformulate, don't paginate.** If the first pass is weak, **rephrase** the
query and re-run nodes+facts (up to 2 more times, accumulating). Stop early
if a rephrase returns essentially the same weak results. Never fetch page 2
just because `has_more` is true — a different phrasing beats pagination.
The identifier pair (rule 7) inverts this: a rephrase leaves the exact route
and cannot stand in for a pair that came back empty — report that as its own
finding, never with the near-miss keys a phrasing returned — and a pair with
`has_more` is paged, because the route matches a term rather than ranking by
meaning.
4. **Summary-first.** Read summaries (node `summary`, lesson `summary`, `search`
`title`) before fetching any full episode body. Full bodies only to cite or
recover a crucial missing detail.
5. **Count caps, not truncation.** No `summary_only` / `max_chars` parameter
exists — bound context with `max_nodes` / `max_facts` / `limit` and by
choosing summary-shaped tools. Never invent a truncation flag.
6. **Bare tool names.** Call `search_memory_nodes` etc. by bare name; the
`mcp__…__` prefix varies per install — use whatever your tool list surfaces.
7. **Identifier lane.** When the question carries an identifier — a prefix of
up to ten characters, a hyphen and digits, whether the user typed it or you
lifted it from the ticket, report or result in hand — call **both**
`search_memory_nodes` and `search_memory_facts` with that identifier as the
**entire** query and no `agent_id`, `center_node_id` or `center_on_user`. Any
extra word, and any of those three, drops the call off the exact route and
onto look-alike keys. One pair per identifier, first, and in addition to the
phrasings — never instead of them. A list of items is not one question: a
skill working through a slice says which of its keys get a pair, and when.
## When to use
- Before any non-trivial task, to surface prior decisions, lessons, and context.
- To answer "have we done this / why did we / what do we know about X / who owns Y".
For **writing** memory see `memory-capture`; for **multi-hop traversal** see
`graph-traversal` (rung 3).
## The search ladder
### Rung 1 — the adaptive first pass (the workhorse)
**Identifier first (rule 7).** If the question carries one, run step 1's two calls
with the bare identifier as the whole query of both, before the phrasings — and
keep the phrasings: nodes return the item and the entities whose summaries name
it, facts return what it relates to, and only a phrasing reaches what was said
around it. An empty pair, or one with `has_more`, is rule 3's case, not a
reason to rephrase. Which shapes fire the exact route, and which spellings of a
key match, are in the tool reference.
1. Run **`search_memory_nodes(query, group_ids, max_nodes≈10)`** and
**`search_memory_facts(query, group_ids, max_facts≈10)`** together on your best
phrasing.
Nodes give entity summaries ("what/who is X"); facts give the relationships
and specific claims ("why / who decided / what's linked"). Most real
questions need **both** — facts frequently carry the actual answer and rank
it higher than the entity summary does.
2. **Judge the results against the question.** If the top few clearly answer it
→ stop and answer. This is the common, cheap case.
3. **If weak** (nothing clearly on-topic) → **reformulate** — different wording,
synonyms, a more specific angle — and re-run nodes+facts. Repeat at most
twice; stop as soon as a rephrase adds nothing new. Accumulate across
phrasings; the best phrasing is often not the first.
4. **Relevance gate before answering:** cite only genuinely on-topic results.
If after reformulating nothing fits, report "no relevant memory found"
rather than stretching a distractor into an answer.
Skip `fetch_lessons_learned` in rung 1 — Lesson entities are also nodes, so
`search_memory_nodes` already surfaces them. Reach for it only when the task is
explicitly about lessons/pitfalls, where its `guidance` / `outcome` structure
is worth a dedicated call.
### Rung 2 — narrow with filters (only if rung 1 left a gap)
Introspect types with `get_available_schemas` / `get_schema` (v2.0 — skip if
hidden), then refine:
- Entities: `search_memory_nodes(entity="<Label>", center_node_id=<id>)`.
- Relationships: `search_memory_facts(edge_type="<TYPE>", center_node_id=<id>)`.
- Time-bound: `search_memory_facts(created_after=…, created_before=…)` —
current-valid is the default; `include_invalidated=true` only for history.
- To scope facts to a person/agent, pivot via `center_node_id` from an
already-scoped node (`search_memory_facts` has no `agent_id`).
### Rung 3 — deep traversal (hand off)
Multi-hop paths and full neighborhoods — `find_path`, `get_node_edges`,
walking edges — belong to the **`graph-traversal` skill**. Cross-reference it;
do not inline traversal here. Those nav tools
are **unbounded** (no pagination, no validity filter) — one more reason to let
that skill own them.
## Summary-first discipline
The three summary surfaces — read these before any episode body:
- **node `summary`** — from `search_memory_nodes`.
- **lesson `summary` / `context_summary`** — from `fetch_lessons_learned`
(`context_summary` is often empty for edge-derived lessons; `summary` is the
dependable field).
- **result `title`** — from the OpenAI-compatible `search`, when present.
Full episode bodies (`get_episode`, `get_episodes`, `get_episodes_for_entity`)
only to quote verbatim or recover a crucial detail. These list endpoints return
the **full body of every item**, so page small (`limit`) and call them **last**.
There is no server-side episode-summary retrieval — summary-first is a
reading discipline, not a tool flag.
## Degradation
Probe with ToolSearch before assuming a tool exists. Visibility is gated in two
**independent** ways — a tool shows only if its **version** ≤ the install's
`ENABLED_TOOL_VERSIONS` _and_ its **tag** is in the `TOOL_PROFILE` (see
`references/tools.md` for the full map).
- **Rung 1 always works:** `search_memory_nodes` and `search_memory_facts` are
`core`-tagged and v1.0 — present under any profile.
- **Hideable:** schema introspection and traversal (`navigation` / `schema`
tags, v2.0) and the `search` / `fetch` pair (v3.0) — each can drop out by
version _or_ by profile (e.g. `agent-lite` omits `navigation` even at v2.0).
If those are hidden: stay on rung 1 and its reformulation loop, skip schema
introspection, drop the `search` titles surface. If the memory server is absent
entirely: state the degradation in one line and proceed — never stall.
## References
- `references/tools.md` — exact per-tool parameters, return shapes, version
tiers, scoping behavior, and the identifier route's contract: which shapes
fire it, which spellings of a key match, which near-miss keys do not, and
what an empty page proves at each offset.
Pliki w paczce (2)
- SKILL.md 8 KB
- references/tools.md 13 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · ibrain-bvba/gutt-claude-code-plugin | 90 | MIT |
| rjmurillo/ai-agents | 100 | MIT |
| rjmurillo/ai-agents | 100 | MIT |
| S3YED/appie-kit | 100 | MIT |
| S3YED/appie-kit | 100 | MIT |
| davila7/claude-code-templates | 95 | MIT |
| 23blocks-OS/ai-maestro-plugins | 95 | MIT |
| doloopz94/virtual-office-muse-ai | 90 | MIT |
| aristoapp/membase | 90 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 967dd975a5b0 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 95pact-memory
Zapisuje kontekst, cele, wnioski i decyzje agentów PACT oraz przeszukuje je semantycznie między sesjami. - 90yggdrasil-memory
Przywołuje wcześniejsze decyzje, rozwiązania i wnioski z poprzednich sesji i projektów, a po rozwiązaniu problemu zapisuje je na przyszłość. - 90sugar-memory
Korzysta z lokalnej pamięci Sugar w Hermes: przypomina wcześniejsze decyzje i błędy przed nowym zadaniem i zapisuje to, co trwałe. - 100recall
Wyszukuje w pamięci wcześniejsze decyzje, uzasadnienia i wzorce, aby stosować je spójnie z poprzednim rozumowaniem. - 100daily-log
Zapisuje sesje pracy, decyzje i ich wyniki oraz wyciągnięte lekcje. - 100using-memory
Przypomina wcześniejsze decyzje i kod z lokalnej pamięci memxt, gdy użytkownik odwołuje się do dawnej pracy, i zapisuje nowe decyzje.