Memind

Memind is an open-source Java memory engine for developers building AI agents that need context across sessions. It turns conversations, documents, tool calls, and resolved tasks into connected memories, while retaining the original material so developers can trace a recalled fact to its source. It's licensed under Apache 2.0.

The system separates user memory, such as preferences and ongoing projects, from agent memory, such as durable instructions, tool experience, and reusable playbooks. It also accepts images, audio, and agent activity timelines. This makes it relevant to coding assistants, personal agents, and chatbots with continuing relationships or tasks.

Memind connects related facts in a memory graph and groups ongoing topics into threads. Its Insight Tree combines individual memories into broader summaries and patterns. Time information helps retrieval distinguish past decisions from current plans or future intentions.

For recall, Memind combines keyword and vector search with graph relationships, threads, and time signals. A deeper search mode uses an LLM to expand queries and check whether the retrieved context is sufficient, with optional reranking. Results can include source references and evidence.

The memory server and Memind UI expose stored content, relationships, insights, and retrieval traces for inspection. Official integrations connect it to Claude Code, Codex, OpenClaw, and Hermes; custom applications can use REST, HTTP MCP, or SDKs for TypeScript, Python, Java, Go, and Rust.

You can self-host the server and UI with Docker Compose, or run them from source. The default setup stores data locally in SQLite and a file vector store. Chat and embedding models use your configured provider endpoint; the example configuration uses OpenRouter, so those model requests leave your machine unless you configure a local endpoint.