Tencent’s Hy-Memory-30B-A3B is a model for deciding what an AI agent should remember. Getting it running locally means deploying a roughly 30-billion-parameter model, connecting it to a memory system and checking whether the resulting stored facts help later answers. Downloading the checkpoint alone does not give every application persistent memory.
The official model card describes about 3 billion active parameters and Apache 2.0 licensing. Its specialization is the write path: choosing facts to retain, representing them precisely and updating them when later evidence changes the situation. That makes it useful to evaluate separately from the model that answers the user’s questions.
Model, framework and app serve different jobs
Tencent’s Hy-Memory 2.0 release page presents three layers. The checkpoint handles memory reading and writing. The Hy-Memory framework manages stored state, updates and recall. MemHub brings a shared memory into supported AI tools. Each can be used separately, so the installation decision should start with the application you actually need.
For an existing Mem0 or MemOS deployment, the interesting change is replacing its general-purpose memory writer. For a new agent, the framework may matter more than the checkpoint. For personal use across tools, start by inspecting MemHub’s integrations rather than assuming a model server automatically intercepts every conversation.
A concrete example shows the distinction. An agent may record that a project uses Python 3.12. If a later message says it has moved to Python 3.13, useful memory must handle that update and retrieve the applicable version. A model can propose the new fact; the application still needs an explicit policy for replacing, retaining or qualifying the older record. This is an illustrative requirement, not an EyesTech test result.
Local serving: budget for the whole model
The official repository lists about 60.1 GB of files and BF16/F32 tensors. The roughly 3 billion active parameters describe conditional computation. They do not mean only 3 billion parameters need storage.
For a rough planning calculation, 30 billion parameters at two bytes each is 60 billion bytes, or approximately 55.9 GiB. That is an illustrative weight-only estimate. Runtime buffers, context state and other processes need additional capacity; it is not a measured peak-memory figure or a guarantee that a particular GPU fits. Quantization changes the arithmetic but also introduces format, runtime and quality questions.
Tencent’s model card identifies vLLM 0.31.0 as its tested serving version and uses the served model name hy-memory. The client example points to an OpenAI-compatible endpoint at http://127.0.0.1:8000/v1, extracts explicit facts into JSON and sets reasoning_effort to no_think through the chat-template options. Those details are worth preserving when reproducing the documented path. We have not run this model locally.
The framework README separately documents the Python package hy-memory at version 2.0.4, with optional storage and reranking integrations. Installing that package and serving the checkpoint are separate operations. Confirm which writer endpoint, storage backend and recall configuration your chosen mode actually uses before calling a deployment local or offline.
What the benchmark lead means
The technical report evaluates fixed 20% benchmark subsamples. The card reports a 55.49 overall average across four memory systems, compared with 54.62 for Claude Opus 5. The difference is 0.87 percentage points, calculated by subtraction. It is not evidence that Hy-Memory wins every task or is a stronger general assistant.
The evaluation averages six metrics within each system and then averages the systems. That weighting matters: changing the mix of coding, shopping, travel or conversational questions changes what the headline average represents. EverMemOS was held out from training, while Hy-Memory, Mem0 and MemOS were used during training. The held-out result provides a useful transfer check within this evaluation’s scope.
Training also explains the emphasis. The report describes on-policy distillation followed by reinforcement learning using downstream answerability, chunk-level feedback and controls against excessive storage. The goal is to improve the usefulness of the written bank, rather than rewarding an indiscriminate transcript dump.
Evaluate the memory bank before connecting real work
A small deployment evaluation can reveal failures that an aggregate score misses. Prepare synthetic sessions with explicit preferences, a later correction, a temporary instruction and two unrelated users. Freeze the answering model and retrieval settings, then change only the writer. Record the stored facts and the final answers, together with write latency and storage volume.
Check whether the old preference survives incorrectly, whether a temporary instruction becomes permanent and whether one user’s facts appear in another user’s recall. For each answer, retain the source message and retrieved memory so a wrong result can be traced to extraction, updating, retrieval or generation. This is a proposed evaluation procedure, not completed testing.
Hy-Memory is most compelling when the writer is a known source of errors or cost in an existing memory pipeline. A production decision needs improvement on your sessions, acceptable serving requirements and inspectable updates. Tencent’s reported benchmark lead is a reason to evaluate that component; the application still has to demonstrate reliable memory over time.
Reporting note: This article was prepared with AI assistance from the linked primary documentation. Calculations are explicitly labeled; no independent model benchmark was performed. The featured image is an original AI-generated editorial illustration.
