Enable coding agents to retain context across sessions
You no longer need to explain the same architecture, re-find the same bugs, or repeat the same preferences. This tool automatically records and compresses agent actions, injecting only the necessary context into the next session. It provides shared memory across a single server for various coding agents, including Claude Code, Cursor, and Codex CLI.
Continuity of memory across sessions
Static files like CLAUDE.md or .cursorrules are limited to 200 lines and quickly become outdated. This tool quietly captures agent actions and compresses them into searchable memory. If you set up JWT authentication in the first session, the agent already knows that src/middleware/auth.ts uses the jose middleware and that jose was chosen over jsonwebtoken for Edge compatibility when you request rate limiting in the second session. The context from previous work continues naturally without separate explanations or copy-pasting.
Search accuracy and token efficiency
It records 95.2% R@5 accuracy on the LongMemEval-S benchmark, which is higher than the 86.2% achieved with BM25 alone. When using local embeddings, the annual cost is $0, and basic functionality is available via BM25-based keyword search without an API key. Hybrid search combines BM25, vector, and graph matching to consider both temporal order and semantic relevance.
Installation and agent connection
Installation begins with the npx -y @agentmemory/agentmemory@latest command in a Node.js 20+ environment. On first run, an interactive setup allows you to select the agent to use and specify an LLM provider or use it without a key. iii-engine is installed automatically on macOS and Linux, but on native Windows, you must use WSL2 or Docker Desktop or manually install a specific version of iii.exe. You can connect more than 20 agents, including Claude Code, Cursor, and Codex CLI, using the agentmemory connect <agent> command, and all agents share the same memory server.
Lifecycle management and real-time viewer
Memory is managed through a four-stage integration, decay, and automatic deletion process. The real-time viewer on port 3113 allows you to replay sessions, playing back prompts, tool calls, and results on a timeline. You can adjust the speed from 0.5x to 4x to track the flow of the session and import Claude Code JSONL transcripts to include existing sessions in search and replay. The import-jsonl command integrates past session records into the current system.
Pre-use considerations
This tool is pinned to iii-engine v0.22.1 and does not work with other engine versions. Windows users must check the limitations of the native environment, with WSL2 being the most stable path. Without a key, vector embeddings are disabled, allowing only keyword search instead of semantic search. To use a local embedding model, set EMBEDDING_PROVIDER=local and allow the initial model download. Data storage follows platform-specific default paths, which can be explicitly specified with the --data-dir option.