Tencent Hunyuan open-sourced Hy Memory 2.0 on October 16–17, and the piece worth your attention isn't the benchmark table. It's the decomposition. The release bundles three things: Hy-Memory-30B-A3B, an Apache 2.0 mixture-of-experts model with roughly 30B total and 3B active parameters, the `hy-memory` SDK, and MemHub, a local-first application. Weights are on Hugging Face under `tencent/Hy-Memory-30B-A3B`, framework and eval code on GitHub under `Tencent-Hunyuan/Hy-Memory`, with a technical report at arXiv 2610.22196. The model is servable through vLLM with an OpenAI-compatible endpoint.
What makes it unusual is scope. It isn't a chat model or a reasoner. It is trained for one step: the memory write — what to retain from a session, how to structure it, and how a new fact supersedes an old one.
The write path has been the unprincipled half
In most production memory stacks, writing is a prompt. You take a session, hand it to the same frontier model serving your users, and ask for extracted facts, a summary, and a dedupe pass. Conflict resolution — the new address contradicts the stored one, the project switched build systems, the user changed their mind in turn 40 — is handled by hand-written heuristics wrapped around that call, or not handled at all until retrieval surfaces two contradictory chunks and the reader model picks one.
Splitting that out has real operational consequences. A 3B-active writer is cheap enough to run asynchronously over every session rather than sampling, and it runs on your own hardware, which matters if your conversation logs are the thing you can't ship to an external API. Tencent reports 72.6% LoCoMo accuracy on chat memory alongside a drop from 1,159 to 703 tokens per question. That second number is the economic argument: better write-time curation means the read path stops paying to haul around raw history. A memory bank is compression, and compression quality is set at write time.
Where the claims are thin
Be skeptical of the headline capability framing. One analysis of the release found the model's average edge over Claude Opus 5 to be 0.87 points, and 0.01 on one harness — effectively parity, not a leap. The reported evaluation spans LoCoMo, LongMemEval, PersonaMem-v2, task benchmarks for travel and shopping, and Code Memory Bench. These are multi-session conversational and repo-QA settings. None of them look like an enterprise knowledge base with per-document ACLs, tenancy boundaries, and sources that mutate underneath the index. Nothing here tells you how a learned write policy behaves when it must *not* remember something.
So: if you're self-hosting coding agents or long-horizon assistants and your memory layer is already a pile of extraction prompts, this is worth a bake-off — the license and the local deployment are the draw, not the scores. If your memory is 200 lines around pgvector and it works, the interesting part is the architecture, not the artifact. Steal the split; you don't have to adopt the model.