agentmemory — память для ИИ-агентов

★ 8.0 · vibe-coding

agentmemory is an open-source persistent memory system for AI coding agents: an MCP server with plugins and hooks that lets an agent remember a project across sessions. It is built on the iii engine. Reach for it when you are tired of re-explaining project layout, when Claude forgets code conventions after a few conversations, when you need to recall decisions from last week, or when several agents should share one knowledge base. It extends Karpathy's LLM Wiki idea with confidence scoring, lifecycle, a knowledge graph and hybrid search. Install with npx -y @agentmemory/agentmemory@latest (Node.js 20+): a wizard wires the agents you choose, including Claude Code (plugin, 12 hooks and MCP), Codex CLI, GitHub Copilot CLI, Cursor, Gemini CLI, OpenCode, Hermes, OpenClaw and any MCP client; separately npx skills add agentmemory -y adds 17 skills. Without keys it uses BM25 search; for local semantic search set EMBEDDING_PROVIDER=local (model Xenova/all-MiniLM-L6-v2). Locally it uses ports 3111, 3112, 3113 and 49134. On native Windows the iii engine is installed by hand; WSL2 or Docker Desktop is easier. It is memory across sessions, not a code generator; it will not fit where Node.js 20 is missing or local ports cannot be opened. Apache-2.0 license.