rohitg00/agentmemory Tops GitHub Trends, Giving AI Coding Agents Persistent Memory

Built on the iii engine, the project requires Node.js 20+; first run interactively selects which agents to connect, provisions config, and starts the memory server. In keyless mode it uses BM25 retrieval with vector embeddings disabled; setting EMBEDDING_PROVIDER=local downloads all-MiniLM-L6-v2 locally for zero-cost semantic recall. It uses fixed ports 3111, 3112, 3113, and 49134, storing state under an agentmemory data directory in the user folder, overridable via --data-dir or an environment variable. For users of Claude Code, Cursor, Codex CLI and similar agents, agentmemory auto-captures sessions via hooks and MCP, compresses them into a searchable memory store, and injects relevant context at the next session start, with no manual copy-paste. The project reports that versus pasting full context (~19.5M tokens/year), this approach uses only ~170K tokens, cutting annual cost from $500 to $10. On LongMemEval-S, agentmemory reaches 95.2% R@5, above pure BM25's 86.2%. It ships 20 adapters covering mainstream coding agents plus a built-in live viewer, making cross-session memory reusable infrastructure for teams running multi-agent workflows.