Knowledge Graph (Anthropic)

This is Anthropic's official memory layer for Claude. It builds a knowledge graph from your conversations, storing entities (people, projects, concepts) and the relationships between them. The graph persists across sessions, so Claude can recall context from weeks ago without you pasting it in again. I've used it for tracking client projects where details accumulate over months. Claude remembers that Project Falcon uses Next.js, that Sarah is the PM, that the staging server had CORS issues last Tuesday. When I ask "What's the status on Falcon?", it pulls from the graph instead of making me repeat myself. It's genuinely useful for long-running work where context matters more than raw compute. The implementation is simple: entities and relations stored locally, queryable via natural language. No vector embeddings, no RAG complexity. Just a graph that Claude can read and write. It works entirely within Claude Desktop and Claude Code, no API keys or external services. Quirks: the graph is only as good as what you explicitly discuss. If you never mention that Sarah reports to Mike, Claude won't infer it. You also can't inspect the graph directly without asking Claude to describe it, which feels odd when debugging. And because it's file-based, moving between machines means manually copying the graph file. This isn't for everyone. If you use Claude for one-off tasks or quick scripts, you won't notice the benefit. But if you're managing ongoing projects, tracking research, or building up domain knowledge over weeks, it changes how Claude works. Instead of context windows, you get something closer to actual memory. Skip it if you context-hop between unrelated tasks or prefer stateless sessions. Install it if you've ever wished Claude remembered last week's conversation without you having to summarise it.