In-depth analysis · About 5 min read
Tencent Hy-Memory 2.0: three layers of memory, with different evidence for each
The October 16 release combines a specialist model, agent framework and MemHub app; its benchmark and internal-business results have distinct scopes.
- Source published
- Published here
- Source checked
Tencent’s official Hy-Memory 2.0 page, dated October 16, presents a memory specialist, a framework and an application that can be used separately or together. The useful question is not simply whether an agent “remembers”: it is what gets retained, how relevant memories are selected and which evaluation supports each claimed improvement. We read the full release, without running its software.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The stack comprises Hy-Memory-30B-A3B, the Hy-Memory framework and MemHub. Chat, working and coding memory are different use cases, not three interchangeable scores.
- The model’s reported 55.49 average versus Opus5’s 54.62 combines six metrics across four systems. Three systems were used in reinforcement learning; EverMemOS was unseen during training.
A specialist, a framework and an app are separate layers
The page identifies Hy-Memory-30B-A3B as a model for reading and writing memories, with approximately 3B active parameters. The framework manages storage, updates and recall through an SDK. MemHub connects memory to the tools people already use. Tencent says all three layers are released openly and may be adopted independently or combined. That distinction matters: replacing a memory model is not the same intervention as adding a framework or installing an application. The page supplies links, but we have not inspected their licenses or tested the deployment requirements.
What changes between chat, working and coding memory
Chat memory organizes events across time and conversations, joins related developments, removes repetition and resolves conflicting information before recall. Working memory tracks goals, settled approaches, intermediate results and reusable experience, checking when that experience applies. Coding memory first identifies the current task and retrieves relevant decisions, still-applicable lessons and previous work. Our interpretation is that the release targets continuity and relevance, not unlimited retention of everything. A useful test includes changed preferences and obsolete decisions: remembering an old fact is insufficient if the system continues to apply it after its conditions have changed.
Framework scores depend on the retrieval setup
Tencent reports 72.6% LoCoMo accuracy, 55.31% on MemoryArena, 64.2% question answering and 35.6% solving on CodeMemoryBench, and 28.3% solving on SWE-ContextBench. Its framework comparison uses DeepSeek V4.1 Flash for memory and GPT-5.6-sol for answering and judging. LoCoMo covers 1,527 non-adversarial questions. Storage occupancy counts generated memory tokens relative to input tokens, excluding raw conversations; the page warns that Simple RAG’s unextracted windows are not directly comparable. These conditions prevent treating either a benchmark rank or a storage ratio as a universal quality, total-disk or operating-cost guarantee.
The specialist-model average is another experiment
The model comparison reports 55.49 for Hy-Memory-30B-A3B and 54.62 for Opus5. It averages six metrics equally across Hy-Memory, Mem0, MemOS and EverMemOS. The first three systems were used during reinforcement learning; EverMemOS is the unseen-system generalization check. Thus the average is evidence about this memory evaluation, not a general ranking of intelligence or proof that every individual metric improves. Nor should it be mixed with the framework scores: those compare memory frameworks under a specified answering setup, while this experiment changes the memory model across systems.
MemHub shares memory only through connected tools
Tencent describes authorized import of historical conversations into local memory, continuing synchronization and plugin-based recall across ten supported tools. The list includes Cursor, Codex, Claude Code, OpenCode and Qwen Code; the page offers CLI and Mac App entry points. This is an integration claim, not confirmation that every tool version or operating system was tested here. Our practical boundary is to inspect what an integration imports, where processing occurs and how access is withdrawn before connecting sensitive work. The phrase local memory alone does not establish that all model calls and synchronization remain offline.
Internal business examples are useful but bounded
Tencent reports an investment-research memory score moving from 64.3 to 79.6, a relative 23.8% increase; it is not an investment return. An iWiki comparison on 251 questions reports 24.7 using all original dialogue versus 36.7 using memory only. Tencent News reports a 91.68 average and 87.8% of cases above 90, but the sample is ten users and 41 complex conversations. These are separate internal evaluations, not population-wide effects. Our recommendation is to reproduce the relevant task with current evidence and a clear acceptance rule before extrapolating to production, privacy or business value.
Why it matters
Teams can evaluate whether remembered decisions and task progress reduce repeated setup in their own agents. A pilot should separate retrieval accuracy, task success and evidence freshness, and confirm authorization before importing conversations across tools.
Limits and availability
These are Tencent’s evaluations, not AISubHub reproductions. We did not read the linked paper, repositories or model license. An open-release claim does not itself verify commercial permissions, security, offline operation, API prices or subscription benefits.
Sources and evidence
Source published: Oct 16, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.