Agentmemory

Agentmemory is an open source persistent memory runtime for AI coding agents. It captures what happens in each coding session and recalls the relevant parts in the next one, running entirely on the local machine with no external database, queue or vector store. It is built for developers who use agents such as Claude Code, GitHub Copilot CLI, Cursor, Gemini CLI, Codex CLI, OpenCode and Hermes and want those agents to stop starting from zero. Twelve automatic hooks record session starts, prompts, tool calls and stops as compressed observations stamped with their origin. Recall combines BM25, vector and knowledge graph scores and reranks them on the device, while keeping superseded versions out of results. With an LLM provider key, consolidation runs at the end of a session to merge duplicates, compress raw observations into semantic memories and decay stale entries, with an audit trail. The full surface is exposed through 54 MCP tools and matching REST endpoints, memory can be scoped per agent, and Claude Code JSONL transcripts can be imported and indexed. The vendor reports 95.2 percent retrieval recall at five on the LongMemEval S benchmark with about 92 percent fewer input tokens than full context replay. Agentmemory is free under the Apache 2.0 license. There is no paid plan; the only optional cost is the LLM provider used for consolidation, and it can run without a key. Setup is a single npx command on Node.js 20 or newer, followed by an interactive wizard that wires in the chosen agents.