openmemind/memind

What it solves

Memind is a memory and context engine designed to solve the problem of AI agents losing context across sessions. It transforms raw data—such as conversations, documents, and tool calls—into structured, evolving memory that allows AI systems to learn from every interaction and reuse experience over the long term.

How it works

Memind processes raw context through a pipeline that converts it into several layers of memory:

• Insight Trees: Distills raw memories into a hierarchy of Leaf, Branch, and Root insights to identify patterns and high-level understanding.
• Memory Graphs: Connects entities, semantic links, and causal signals to recover related context that simple vector searches might miss.
• Memory Threads: Groups related items into durable timelines to help agents track unfinished work and reuse task history.
• Multi-layer Retrieval: Combines vector search, BM25 keyword search, temporal signals, and deep retrieval (query expansion and reranking) to recall the most relevant context.

Who it’s for

It is intended for developers building AI systems that require long-term memory, including:

• Coding agents needing project context and tool experience.
• Personal agents tracking user preferences and workflows.
• Chatbots and companions maintaining user profiles and life events.
• Workflow agents remembering directives and playbooks.

Highlights

• SOTA Performance: Ranks #1 on LoCoMo, LongMemEval, and PersonaMem benchmarks.
• Java-Native: The first state-of-the-art memory engine built natively in Java.
• Dual Memory Scopes: Separates user memory (profiles, preferences) from agent memory (directives, tool experience).
• Broad Integration: Supports REST APIs, SDKs (TypeScript, Python, Java, Go, Rust), and the Model Context Protocol (MCP).
• Versatile Ingestion: Can ingest text, images, audio, and agent timelines.