Hy Memory

What is Hy Memory?

Hy Memory is an open-source Python SDK and memory engine for self-evolving LLM agents. It uses a dual-process architecture to organize information into hierarchical long-term memory for chat, two-layer fact storage for working tasks, and version-tracking chains for coding workflows. Specialized pipelines manage how agents store and retrieve knowledge.

MemHub is a companion local-first application built on Hy Memory. It aggregates conversation history and coding traces from authorized AI tools like Claude Code, Cursor, and Qwen Code into a single runtime and local data directory. This setup allows users to maintain controlled, cross-application memory recall without relying on cloud services. The project provides both a command-line interface and a desktop application for macOS and Linux.

The framework supports multiple memory modes, automatic retry mechanisms for memory writes, and safe switching between embedding models. Quantized model versions are available for local deployment. Research backing the system was accepted at EMNLP 2026, focusing on reinforcement learning for long-term memory writing.

Hy Memory features

• Dual-process cognitive memory architecture for agents
• Hierarchical long-term memory for chat and working tasks
• Version-evolution tracking for coding workflows
• Local-first desktop and CLI application called MemHub
• Aggregates data from authorized AI coding assistants
• Supports multiple memory modes and safe embedding switching
• Automatic retry logic for memory write operations

Who uses Hy Memory and for what

• Building autonomous coding agents that retain project context across sessions
• Maintaining personal knowledge bases for chat-based AI assistants
• Unifying conversation history from multiple AI tools into one local directory
• Testing long-horizon agent memory benchmarks with the CMB dataset
• Deploying quantized memory models locally for privacy-focused applications

Hy Memory FAQ

Comments (0)

Comments and upvotes open when this product launches on ProductPerch.