Arcadedb Claude

arcadedb-claude

Persistent graph-based memory and code intelligence for AI coding agents (Claude Code, Cursor, Aider, custom Anthropic SDK / OpenAI SDK agents). Multi-repo, multi-project, locally hosted, open source. Powered by [ArcadeDB](https://arcadedb.com).

[](https://www.npmjs.com/package/arcadedb-claude-skills) [](https://opensource.org/licenses/MIT)

The problem

AI coding agents are stateless. Each session starts cold:

• They cannot recall decisions you made together last week (why you chose Tailwind over Bootstrap, why that abstraction was rolled back).
• They re-read whole files to answer structural questions like "what calls this function?" or "what depends on this module?" — burning context tokens on work the codebase's own structure already encodes.
• They cannot reason across multiple repos at once. Your monorepo, your standalone backend, and your client codebases live in disjoint mental models.
• Per-project facts (which database does this project use, who paged about it last quarter) live in scattered CLAUDE.md files, Slack threads, and your head.

What's needed is a **persistent graph** the agent can query before answering — one that holds both code structure (imports, calls, contains) and accumulated memory (decisions, insights, sessions) — and that spans every repo you work on.

The solution

`arcadedb-claude` ships as one npm package, [`arcadedb-claude-skills`](./packages/claude-skills), that gives Claude Code (and any other Anthropic SDK or OpenAI SDK agent that can shell out to a CLI) exactly that:

Component	What it does
Plugin (hooks + commands)	Auto-injects per-project graph context on every session start. Slash commands for recording decisions, querying the graph, indexing, and status
src/agent-memory + `arcadedb-memo