Intel / article

Tencent open-sources Hy Memory 2.0 with RL-trained memory writer

Source: github.com
Author: Tencent Hy
Date: Published Oct 16, 2026

Why it matters

A small (~3B active) model can replace the LLM in Mem0, MemOS and Hy-Memory write paths, cutting memory-writing cost. The framework also adds coding-memory diff chains and MemHub cross-tool memory for Claude Code, Cursor and Codex.

Key takeaways · AI-distilled

Hy Memory routes memory by scene: chat uses an L0-L7 hierarchy from raw dialogue through facts, summaries and schemas to expiring intentions, while working and coding store one session fact plus reusable experiences.
Coding extraction records file, function and branch coordinates instead of pasting code, so Hy Memory adds a diff chain that stores every revision's real diff with zero calls, preserving versions that were later reverted.
Three modes trade cost for depth: lite does raw writes and searches with no LLM calls, pro adds LLM extraction and conflict reconciliation, and ultra adds System 2 schema induction on an always-on graph store.
When no scene is given, writes without tool calls go straight to chat memory with no classification call, and only writes containing tool calls are classified once into chat, working or coding.
Tencent also released Hy-Memory-30B-A3B, a model trained with attributed reinforcement learning specifically for long-term memory extraction and conflict reconciliation, with a technical report on arXiv.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

Read the source github.com