Graph Memory

该 Skill 提供图谱实体搜索、图谱统计和关系遍历三类函数，用于按自然语言或参数化方式查询智能体积累的会话记忆。结果以 Markdown 文本返回，包含实体类型、提及次数、关系路径与基础统计。

space-cadet

热度 15.83

最近更新 2026-07-23 14:18:11

分类：知识依赖：代码执行 / 读写文件应用场景：研究 / 笔记整理

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add graph-memory

Metadata

name: graph-memory
description: Query the agent's knowledge graph for entities, relationships, and statistics. Enables natural language exploration of accumulated session memory without shell commands.
metadata: {"openclaw":{"emoji":"🕸️","requires":{"modules":["better-sqlite3"]},"install":[{"id":"npm-better-sqlite3","kind":"npm","package":"better-sqlite3","label":"Install better-sqlite3 (npm)"}]}}

Graph Memory

Query the agent's persistent knowledge graph — entities extracted from sessions, their relationships, and graph-wide statistics.

The graph is built from session history and contains people, projects, tools, concepts, files, errors, and more. Use this skill when the user asks about "what do you know about X", "how is X connected to Y", or when you need context about past work.

Functions

`graph_search(query, options)`

Search the graph for entities matching a text query.

Parameters:

query (string, required) — Text to search for in entity names and canonical names
options (object, optional):
- type (string) — Filter by entity type (e.g. person, project, tool, concept, file)
- deep (boolean) — Include relationship neighbors for each result (default: false)
- limit (number) — Max results to return (default: 20)

Returns: Markdown string with matching entities, their types, mention counts, and optional neighbor lists.

Example:

graph_search("chimera", { type: "project", deep: true })

Error cases:

Graph DB missing → "Graph memory database not found at <path>"
No results → "No entities matching '<query>' found in graph memory."
Runtime error → "Graph query error: <message>"

`graph_stats()`

Return high-level statistics about the graph.

Parameters: None

Returns: Markdown string with:

Total entity count
Total relationship count
Sessions processed count
Entity type breakdown
Top 10 entities by mention count

Example:

graph_stats()

Error cases:

Graph DB missing → "Graph memory database not found at <path>"
Runtime error → "Graph query error: <message>"

`graph_related(entity, depth)`

Traverse the graph from a starting entity, showing connected entities up to N hops away.

Parameters:

entity (string, required) — Exact or approximate entity name to start from
depth (number, optional) — Max traversal depth (default: 2, max: 5)

Returns: Markdown relationship tree showing paths from the starting entity to connected entities, with relationship types and entity details.

Example:

graph_related("deepak", 2)

Error cases:

Graph DB missing → "Graph memory database not found at <path>"
Entity not found → "Entity '<entity>' not found in graph memory."
No relationships → "No relationships found for this entity in graph memory."
Runtime error → "Graph query error: <message>"

Entity Types

Common types in the graph:

Type	Description
`person`	People mentioned in sessions
`project`	Projects, repositories, systems
`tool`	Software tools, libraries, CLIs
`concept`	Ideas, theories, patterns
`file`	Specific files referenced
`error`	Error types or failure modes
`topic`	Broad subject areas
`institution`	Organizations, universities
`research_paper`	Academic papers

Notes

The graph database lives at ~/.openclaw/workspace/.openclaw_memory/graph.db
It is populated by a background queue worker that processes session history
Not all sessions have been processed; extraction quality varies
Relationship density is highest for project/file connections; person-to-project links are sparse
Use graph_search with deep: true to see immediate neighbors without full traversal