memory-graph

Memory Graph

Prerequisites

First ensure memory-graph is executable:

• Already ran {Skill}/assets/install.sh
• Execute memory-graph -v can output the version number

Actions

link

Look up the memory graph according to the user's request, and describe to the user your understanding based on historical memories, as well as how it will affect the current task

memory-graph search "<关键词1[空格]关键词2 ...>" [--nodes 5]

patch/apply

Used when the user requests consolidating memories.

• During patch: first review the session history, focus onuser input information
• During apply: first understand the content provided by the user or the user's request
• Then execute the following steps

1. Extract reusable conclusions, evidence, rules, actions, experience, or assets
2. For the user's convenience in editing, first generate/tmp/memory-apply.yaml, provide the user with a clickable preview link
3. Let the user view, edit, or give an optimization direction
4. After receiving the confirmation write instruction, then write to the memory graph
5. Write to memory graph: memory-graph apply /tmp/memory-apply.yaml

Example:

summary: 本轮会话摘要。
nodes:
  - id: note:worktree-reuse-node-modules
    type: Note
    title: worktree 复用 node_modules
    summary: >
      master 已有 node_modules 时，新 worktree 应软链复用，
      避免重复安装依赖。
    tags: [worktree, node_modules, symlink]
edges: []

look

Used when the user needs to view, clean up, or manage memories

memory-graph look

• Supports: overview, search, node details, update/delete nodes, delete relationships

Rules

• Storage always uses global ~/.memory-graph
• nodes only use Concept、Note、Basis、Asset、Rule、Skill、Action
• nodes can write tags, used for lightweight indexing and retrieval; tags are not built into nodes
• Prioritize consolidating reusable conclusions, do not write running accounts, do not write entire chat transcripts as nodes
• patch uses YAML by default, fields remain summary / nodes / edges, temporary files are always written to /tmp/memory-apply.yaml
• Without user confirmation, do not execute apply

Graph model

Nodes

Type	Definition and purpose
Concept	Concepts, modules, project names, tools, commands, person names...
Basis	Raw evidence: user instructions, chat records, code snippets, errors, terminal output...
Note	Stable conclusions, judgments, experience, derived by Basis induction
Asset	Files, directories, commands, configurations, interfaces, functions within the project...
Rule	Rules, user preferences, behavioral constraints...
Skill	Tool invocation experience, operation steps, workflows...
Action	Specific to execution operations, but must include a verb and an object

Edges

Edge	Meaning	Domain (source)	Range (target)
DESCRIBES	Conclusion/rule describes object characteristics	Note, Rule	Concept, Asset
BASED_ON	Conclusion is based on raw evidence	Note	Basis
APPLIES_TO	Rule/skill applicable object	Rule, Skill	Concept, Asset
REQUIRES	Operation/skill depends on a certain condition or artifact	Action, Skill	Asset, Rule, Concept
PRODUCED	Operation produces an artifact or raw output	Action	Asset, Basis
NEXT	Next step in the process	Skill, Action	Skill, Action
USES	Usage relationship	Action, Skill, Concept, Asset	Concept, Asset
CONTAINS	Hierarchical containment	Concept, Asset	Concept, Asset
CAUSES	Causal relationship	Action, Basis	Basis
RELATES_TO	Fuzzy association, used for diary-style management	Any node	Any node