Tencent's Hunyuan team has open-sourced "Hy Memory 2.0," which includes a dedicated memory-writing model, Hy-Memory-30B-A3B, a programmable memory framework, the hy-memory SDK, and an end-user desktop tool, MemHub, forming a complete "memory infrastructure."
Compared with the mainstream approach of wrapping a memory framework (such as Mem0 or MemOS) around a general-purpose large model (such as Claude Opus), Hy Memory 2.0 takes a different route: it replaces the general-purpose large model with a small, specialized model for memory writing, which is both cheap and able to work across frameworks.
The Three Components of Hy Memory 2.0
Model: Hy-Memory-30B-A3B. It is a 30B-total-parameter MoE architecture with only about 3B activated parameters (meaning inference cost is close to that of a 3B model). Training happens in two steps: first, "On-Policy Distillation" lets the student model learn each framework's write protocol on the memory states it generates itself, rather than simply imitating teacher outputs; second, "attributed memory reinforcement learning" directly optimizes downstream performance, using per-chunk probes to locate contributions and carrying the memory store it wrote itself across context windows, keeping training consistent with real deployment scenarios.
Framework: hy-memory SDK. This is the most valuable part from an engineering standpoint. It designs different memory forms for three types of scenarios:
· Chat memory uses an L1–L7 hierarchical structure (raw conversation → atomic facts → session summaries → mental models → forward-looking intentions), offered in lite / pro / ultra tiers, with the lowest tier making zero LLM calls and relying purely on vector retrieval;
· Working memory distinguishes between "session facts" and "reusable experience" layers, and experience writes must pass a value threshold to prevent the memory store from bloating;
· Coding memory records only code coordinates (file/function/branch) rather than pasting the code itself, and attaches the actual diff of each revision to facts via a "decorator" approach, forming a version evolution chain. This design is very effective at controlling noise in coding scenarios.
Storage backends are pluggable (ChromaDB/Qdrant/FAISS + Kuzu/Neo4j + SQLite), and any OpenAI-compatible endpoint can run it. It is ready to use via pip install hy-memory.
MemHub: a local-first CLI and Mac desktop app that unifies the session histories of tools such as Claude Code, Cursor, Codex, and Qwen Code into a single memory runtime, enabling "preferences learned in Cursor can also be used by Claude Code." This is the most intuitive entry point for individual users.
How do the benchmark results actually look?
Across six metrics macro-averaged over four harnesses (Mem0, Hy-Mem, MemOS, and EverMemOS, which did not participate in training), this model ranks first with 55.49, versus 54.62 for Claude Opus 5. The lead is under 1 point, and the data comes from the official technical report (based on a fixed 20% benchmark subsample), so these are self-reported results that still await independent reproduction.
Two points are more convincing: first, it still leads on the held-out framework EverMemOS, which it has never seen (56.89 vs 54.04), demonstrating generalization rather than overfitting to specific protocols; second, compared with its same-architecture base model Qwen3-30B-A3B (44.37), it improves by about 11 points, showing that the training method itself is effective.
Using a model with about 3B activated parameters to approach or even exceed flagship general-purpose models, the core selling point is cost-effectiveness. Memory writing is a high-frequency call, so cost differences get amplified.
SDK & MemHub: https://github.com/Tencent-Hunyuan/Hy-Memory
Model: https://huggingface.co/tencent/Hy-Memory-30B-A3B