topic-memory-graph

中文说明

TypeScript library that builds a topic graph from chat text (with simple decay, associations, and optional mood/trail signals) and renders a Markdown section suitable for injection into AGENTS.md or similar agent context files.

Extracted from the DidClaw client (LCLAW monorepo) for reuse and standalone publishing. Zero runtime dependencies.

Screenshots

The DidClaw desktop app includes a Cognitive Map (Experimental) UI on top of the same graph data: a force-directed network and a structured panel with hot topics and associations.

Install

pnpm add topic-memory-graph
# or: npm install topic-memory-graph

Quick usage

import {
  emptyGraph,
  updateGraph,
  applyDecay,
  generateMemorySection,
  shouldInjectMemory,
  injectMemorySection,
  DEFAULT_MARKERS,
  DEFAULT_INJECT_INTERVAL_RUNS,
} from "topic-memory-graph";

let graph = emptyGraph();
graph = updateGraph(graph, "user message", "assistant reply");

const md = generateMemorySection(graph, { attribution: "MyApp" });
const agentsBody = injectMemorySection(existingAgentsMd, md, DEFAULT_MARKERS);

Persist graph as JSON however you like (filesystem, Tauri, IndexedDB). See integrations/tauri-reference.md for notes aligned with the DidClaw desktop app.

API highlights

Export	Role
extractTopics, updateGraph, applyDecay	Core graph maintenance
generateMemorySection(graph, { attribution? })	Markdown for agents
shouldInjectMemory(graph, runsSinceLastInject, minRuns?)	When to re-inject
injectMemorySection(md, content, markers?)	Safe replace/append via HTML comments
DEFAULT_MARKERS, DIDCLAW_PHEROMONE_MARKERS	Marker pairs for injection

Graph shape

PheromoneGraph is versioned JSON (GRAPH_SCHEMA_VERSION). Hosts should preserve unknown fields if they merge with future library versions.

Emotion modes

A / B / C / N are heuristic labels from regex signals on user text (e.g. narrow focus vs expansive vs low-energy). They are not clinical sentiment scores.

License

MIT. (The wider LCLAW monorepo may use other licenses for unrelated code.)

Developing this package

cd packages/cognitive-memory-graph
pnpm install
pnpm test
pnpm build

Standalone clone: git clone https://github.com/didclawapp-ai/topic-memory-graph.git. In the LCLAW monorepo the same sources currently live under packages/cognitive-memory-graph (rename locally if you prefer to match this repo name).