Tencent Ships Hy Memory so AI Agents Remember Everything Across Sessions

Tencent’s Hunyuan team has released Hy Memory, an open-source framework for LLM agents that retain information across sessions, tools, and projects. The project uses separate memory pipelines for conversation, knowledge work, and coding, alongside a local desktop and CLI application called MemHub.

LLM APIs do not retain state between requests, so agent applications must store useful history and retrieve it when needed. A shared vector index can mix personal preferences, task outcomes, and source-code changes despite their different retrieval requirements. Hy Memory assigns each category its own structure and lifecycle.

The release bundles the following components:

• A Python SDK published as hy-memory==2.0.4.
• A 30-billion-parameter Mixture-of-Experts memory-writing model, including FP8 and GGUF variants.
• MemHub integrations for Claude Code, Cursor, Codex, OpenCode, and Qwen Code.
• Two research papers and the CMB coding-memory benchmark.
• Vector and graph storage integrations for local and service-backed deployments.

The release materials identify the code and model artifacts as Apache-2.0 licensed.

One job, one memory shape

Hy Memory calls its task-specific configurations “scenes.” Automatic routing selects among three scenes, and applications can specify one directly when the task is known.

• Chat uses an eight-layer hierarchy, labeled L0 through L7. It turns raw dialogue into atomic facts, summaries, learned patterns called mental schemas, and prospective intentions such as plans or reminders. Intentions include expiration times and become ordinary facts after they expire.
• Working uses two layers. Each session produces one fact describing the completed work, plus reusable experiences that meet the system’s value threshold.
• Coding extends the working schema with a version-evolution chain that tracks revisions to each function by branch.

The separation gives each retrieval path information suited to its task. Chat searches can surface preferences and plans, working-memory searches can recover prior outcomes, and coding searches can reconstruct how an implementation changed.