cognee:recall — скилл для Claude Code: поиск по графу знаний
Скилл cognee:recall для Claude Code ищет сущности, решения, события и связи в графе знаний Cognee. Использует семантический, графовый и лексический поиск, чтобы добавить найденное в контекст агента.
MIT ★ 328 обновлён 2026-10-29
npx skills add coco-research/coco --skill cognee:recall -a claude-codeЧто делает cognee:recall
Скилл запускает поиск по базе знаний Cognee тремя способами: семантический (по смыслу), графовый (по связям между сущностями) и лексический (по ключевым словам). Возвращает решения, задачи, сущности и события с оценкой релевантности, затем форматирует их как контекст для агента — тот может опираться на прошлые решения и зависимости между модулями при ответе.
Внутри три слэш-команды: /cognee-recall:search для обычного поиска, /cognee-recall:context для инъекции результатов в системный промпт и /cognee-recall:graph для обхода соседей сущности в графе. Тип поиска выбирается автоматически (FEELING_LUCKY) или задаётся вручную — GRAPH_COMPLETION для запросов вроде «кто владеет X», TEMPORAL для поиска по времени, CODING_RULES для кода.
Для работы нужен запущенный экземпляр Cognee API (по умолчанию http://localhost:8000). Адрес задаётся через переменную окружения COGNEE_BASE_URL. Поиск может идти по одному датасету проекта или сразу по нескольким — скилл достаёт список через /api/v1/datasets.
Какие задачи решает
- найти прошлые архитектурные решения по теме
- понять зависимости между модулями и сервисами
- вспомнить, кто отвечает за задачу или компонент
- собрать контекст из нескольких проектов перед ответом
- построить граф связей вокруг сущности
Примеры запросов агенту
- Найди в графе знаний всё про аутентификацию
- Покажи связи для сущности Auth Service
- Вспомни решения по rate limiting из всех проектов
Как установить cognee:recall
1Через Skills CLI (Vercel)
npx skills add coco-research/coco --skill cognee:recall -a claude-code2Вручную: положить файлы
~/.claude/skills/cognee-recall/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/cognee-recall/
1Через Skills CLI (Vercel)
npx skills add coco-research/coco --skill cognee:recall -a codex2Вручную: положить файлы
.agents/skills/cognee-recall/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add coco-research/coco --skill cognee:recall -a cursor2Вручную: положить файлы
.cursor/skills/cognee-recall/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install coco-research/coco systems/cognee/skills/cognee-recall2Вручную: положить файлы
.github/skills/cognee-recall/SKILL.md1Через Skills CLI (Vercel)
npx skills add coco-research/coco --skill cognee:recall -a gemini-cli2Командой платформы
gemini skills install https://github.com/coco-research/coco --path systems/cognee/skills/cognee-recallВручную: положить файлы
.devin/skills/cognee-recall/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о cognee:recall
Как установить cognee:recall в Claude Code?
Выполни npx skills add coco-research/coco --skill cognee-recall -a claude-code, затем проверь ~/.claude/skills/cognee-recall/SKILL.md.
Что делает cognee:recall?
Скилл ищет сущности, решения, задачи и события в графе знаний Cognee через семантический, графовый и лексический поиск, затем подставляет результаты в контекст агента для принятия решений.
Какой API нужен для работы cognee:recall?
Нужен запущенный экземпляр Cognee API, по умолчанию на http://localhost:8000. Адрес задаётся через переменную COGNEE_BASE_URL.
Исходный текст
SKILL.md · MIT
---
name: cognee:recall
description: "Use when the user says 'cognee recall', 'search memory', 'what do we know about', 'find related', or 'graph search'. Runs semantic, graph, and lexical search across the Cognee knowledge graph and injects results as agent context."
---
# /cognee-recall — Search & Recall from Knowledge Graph
Semantic, graph-traversal, and lexical search across Cognee's knowledge graph. Finds entities, decisions, events, and their relationships — then injects relevant results as agent context for informed decision-making.
## Quick Reference
```bash
COGNEE="${COGNEE_BASE_URL:-http://localhost:8000}"
DATASET="my-project"
# Semantic search (auto-selects best strategy)
curl -s -X POST "$COGNEE/api/v1/search" \
-H "Content-Type: application/json" \
-d '{"query": "authentication decisions", "datasets": ["my-project"], "search_type": "FEELING_LUCKY", "top_k": 10}' | jq .
# Graph completion search (relationship-aware)
curl -s -X POST "$COGNEE/api/v1/search" \
-H "Content-Type: application/json" \
-d '{"query": "who reports to Alice", "datasets": ["my-project"], "search_type": "GRAPH_COMPLETION", "top_k": 10}' | jq .
# Recall with context injection (adds system prompt)
curl -s -X POST "$COGNEE/api/v1/recall" \
-H "Content-Type: application/json" \
-d '{"query": "rate limiting", "datasets": ["my-project"], "top_k": 10, "only_context": true}' | jq .
# Cross-project search
curl -s -X POST "$COGNEE/api/v1/search" \
-H "Content-Type: application/json" \
-d '{"query": "auth decisions", "datasets": ["project-a", "project-b", "project-c"], "search_type": "FEELING_LUCKY"}' | jq .
```
## Search Types
Cognee supports multiple search strategies. Use `FEELING_LUCKY` for auto-selection (recommended), or specify one:
| Type | Best for |
|------|---------|
| `FEELING_LUCKY` | Auto-selects best strategy (default, recommended) |
| `GRAPH_COMPLETION` | Relationship-heavy queries ("who owns X", "what depends on Y") |
| `GRAPH_COMPLETION_COT` | Complex reasoning with chain-of-thought |
| `GRAPH_COMPLETION_CONTEXT_EXTENSION` | Expanding context around a node |
| `GRAPH_SUMMARY_COMPLETION` | Summarization of graph neighborhood |
| `RAG_COMPLETION` | Retrieval-augmented generation |
| `TRIPLET_COMPLETION` | Entity-relationship-entity patterns |
| `CHUNKS` | Raw chunk retrieval |
| `CHUNKS_LEXICAL` | Keyword/lexical matching |
| `SUMMARIES` | Pre-computed summaries |
| `NATURAL_LANGUAGE` | Free-form natural language queries |
| `TEMPORAL` | Time-based queries |
| `CODING_RULES` | Code-specific patterns |
## Sub-commands
### /cognee-recall search — Run a semantic/graph search
**Procedure:**
1. Ask the user what they're looking for (or use the provided query).
2. Determine the best search type based on the query:
- Queries about relationships → `GRAPH_COMPLETION`
- General "what do we know" → `FEELING_LUCKY`
- Time-based ("last week", "in Q2") → `TEMPORAL` if available
- Code-related → `CODING_RULES`
3. Detect which datasets to search:
- Default: current project dataset
- If user mentions another project: include its dataset
- Use `/cognee status` to list available datasets
4. Execute search:
```bash
COGNEE="${COGNEE_BASE_URL:-http://localhost:8000}"
curl -s -X POST "$COGNEE/api/v1/search" \
-H "Content-Type: application/json" \
-d "{
\"query\": \"$QUERY\",
\"datasets\": $DATASETS_JSON,
\"search_type\": \"$SEARCH_TYPE\",
\"top_k\": $TOP_K
}" | jq .
```
5. Present results:
```
COGNEE RECALL — "$QUERY"
==================================================
Found N results across M datasets
[1] DECISION: Use JWT for API auth (2026-06-30)
Context: Stateless, works with existing infra
Dataset: my-project | Score: 0.94
[2] ENTITY: Auth Service — depends_on → PlatformHub
Description: Authentication and authorization service
Dataset: my-project | Score: 0.87
[3] TASK: Set up JWT middleware (open, priority 1)
Assigned to: Alice Chen
Dataset: my-project | Score: 0.82
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
```
### /cognee-recall context — Inject results as agent context
Same as search, but formats results for direct injection into the agent's context window. Use this before making architectural decisions or when context from past sessions is needed.
**Procedure:**
1. Run search as above.
2. Format results as a compact context block:
```
[COGNEE CONTEXT INJECTION — {timestamp}]
Query: "{original_query}"
Dataset(s): {dataset_names}
Relevant knowledge:
• DECISION ({date}): {text} — {context} [relevance: {score}]
• ENTITY: {name} ({type}) — {description} [relevance: {score}]
• TASK: {text} ({status}) — assigned to {assignee} [relevance: {score}]
• EVENT: {title} ({date}, {type}) — {summary} [relevance: {score}]
Use this context to inform your response. Cite specific decisions and entities where relevant.
```
3. The agent then uses this context transparently in its reasoning.
### /cognee-recall graph — Explore entity neighborhood
Explore the graph around a specific entity to understand its relationships.
```bash
COGNEE="${COGNEE_BASE_URL:-http://localhost:8000}"
# First, get the dataset ID
DATASET_ID=$(curl -s "$COGNEE/api/v1/datasets" | jq -r '.[] | select(.name=="my-project") | .id')
# Get the full graph
curl -s "$COGNEE/api/v1/datasets/$DATASET_ID/graph" | jq .
```
Present as a relationship map:
```
ENTITY GRAPH — "Auth Service" (my-project)
==================================================
┌──────────────────┐
│ Auth Service │
│ (system) │
└───┬──────────┬───┘
│ │
depends_on │ │ owns
│ │
┌──────▼──┐ ┌───▼──────────┐
│Platform │ │ JWT Middleware│
│Hub │ │ (module) │
│(module) │ └───────────────┘
└─────────┘
Related decisions:
• Use JWT for API auth (2026-06-30) — relates to Auth Service
Related tasks:
• Set up JWT middleware (open) — assigned to Alice Chen
• Update API docs (open) — assigned to unassigned
```
### /cognee-recall cross-project — Search across multiple projects
Search for related knowledge across all available datasets.
```bash
COGNEE="${COGNEE_BASE_URL:-http://localhost:8000}"
# Get all dataset names
DATASETS=$(curl -s "$COGNEE/api/v1/datasets" | jq -r '[.[].name] | join(",")')
# Cross-project search
curl -s -X POST "$COGNEE/api/v1/search" \
-H "Content-Type: application/json" \
-d "{
\"query\": \"$QUERY\",
\"datasets\": [$DATASETS_JSON],
\"search_type\": \"FEELING_LUCKY\",
\"top_k\": 15
}" | jq .
```
Present results grouped by dataset:
```
CROSS-PROJECT RECALL — "$QUERY"
==================================================
Found N results across M datasets
my-project (5 results):
[1] DECISION: Use JWT for API auth — 0.94
[2] ENTITY: Auth Service — 0.87
...
e-and-c (3 results):
[1] DECISION: Stakeholder role for external users — 0.91
[2] ENTITY: External Review System — 0.84
...
optimize (2 results):
[1] DECISION: Migration to pgvector — 0.78
...
```
## Behavior Rules
### Proactive recall triggers
Automatically run `/cognee-recall search` when:
- User asks "what do we know about X" or "what did we decide about Y"
- User starts a new task that references past work ("continue working on auth")
- User makes a decision that might conflict with past decisions
- Before architectural or design discussions that would benefit from context
### Relevance threshold
Only show results with relevance score > 0.5 by default. If fewer than 3 results exceed threshold, tell the user: "Only {N} low-relevance results found. Try a broader query."
### Context injection discipline
- Inject context **before** reasoning about a decision, not after.
- Cite specific decisions/entities from the graph in responses: "Based on the June 30 decision to use JWT for API auth [Cognee]..."
- Never fabricate — if the graph has no relevant data, say so.
### Session-aware recall
If Cognee supports sessions, tag recalls with a session ID for better multi-turn context:
```bash
SESSION_ID="coco-$(date +%s)"
curl -s -X POST "$COGNEE/api/v1/search" \
-H "Content-Type: application/json" \
-d "{
\"query\": \"$QUERY\",
\"datasets\": [\"$DATASET\"],
\"search_type\": \"FEELING_LUCKY\",
\"top_k\": 10
}" | jq .
```
## Fallback Behavior
If Cognee is unreachable:
1. Check if Brain is available: look for `project_brain.db` in current or parent directories
2. If available: "Cognee is not running. Using Brain for this query instead." Run equivalent `/brain context --search "query"`
3. If neither available: "No memory backend available. Start Cognee with `cognee server start` or initialize Brain with `/brain init`."
## Prerequisites
Cognee must be running and the project dataset must exist (created via `/cognee init`).
```bash
# Verify
curl http://localhost:8000/health
curl http://localhost:8000/api/v1/datasets | jq '.[].name'
```
© coco-research · MIT · coco-research/coco