Log in Sign up
News UCLA's PhoneBot Is a $400 Humanoid That Runs on an Old Android Phone's Camera, Sensors and Chip AI Agents Tencent Open-Sources Hy Memory 2.0: a 3B-Active Memory Model That Edges Claude Opus 5 on Its Own Benchmarks Security World Launches a Banking World ID Suite, Betting 18 Million Proofs of Human Can Slow Deepfake Fraud Business Japan Clears 3nm at TSMC's Second Kumamoto Fab: 6,000 Wafers a Month, First Shipments in Q4 2028 Business Gartner Says 80 Percent of Neoclouds Will Stop Operating Independently by 2030, and Executives Reportedly Expect Only 7 to 10 of Hundreds to Survive News China Puts Its First Homegrown 100,000-Card AI Supercluster on Display as Sugon Publishes Ten Results From Dawning 8000 AI Agents Google Grounds AI Agents in Wet-Lab Robots: a Digital Twin Lifted Aptamer Hit Rates to 42 Percent in Three Design Rounds Opinion Taste-Bench From Microsoft and CityU Shows AI Agents Pick the Right Long-Term Move Only 60 Percent of the Time News UCLA's PhoneBot Is a $400 Humanoid That Runs on an Old Android Phone's Camera, Sensors and Chip AI Agents Tencent Open-Sources Hy Memory 2.0: a 3B-Active Memory Model That Edges Claude Opus 5 on Its Own Benchmarks Security World Launches a Banking World ID Suite, Betting 18 Million Proofs of Human Can Slow Deepfake Fraud Business Japan Clears 3nm at TSMC's Second Kumamoto Fab: 6,000 Wafers a Month, First Shipments in Q4 2028 Business Gartner Says 80 Percent of Neoclouds Will Stop Operating Independently by 2030, and Executives Reportedly Expect Only 7 to 10 of Hundreds to Survive News China Puts Its First Homegrown 100,000-Card AI Supercluster on Display as Sugon Publishes Ten Results From Dawning 8000 AI Agents Google Grounds AI Agents in Wet-Lab Robots: a Digital Twin Lifted Aptamer Hit Rates to 42 Percent in Three Design Rounds Opinion Taste-Bench From Microsoft and CityU Shows AI Agents Pick the Right Long-Term Move Only 60 Percent of the Time

Tencent Open-Sources Hy Memory 2.0: a 3B-Active Memory Model That Edges Claude Opus 5 on Its Own Benchmarks

AI Agents · By: AIReport Admin · 2 hours ago · 👁 7
Tencent Open-Sources Hy Memory 2.0: a 3B-Active Memory Model That Edges Claude Opus 5 on Its Own Benchmarks

Tencent's Hunyuan team has open-sourced all three layers of Hy Memory 2.0, an agent memory stack: the Hy-Memory-30B-A3B model with about 3 billion active parameters, the hy-memory SDK and the MemHub local app, all under Apache 2.0. On the company's fixed 20 percent benchmark subsample the memory model scores a 55.49 macro average, slightly above Claude Opus 5 at 54.62, though those figures are self-reported.

Tencent's Hunyuan team has released version 2.0 of Hy Memory, its memory framework for AI agents, and open-sourced the entire stack: a dedicated memory model, the framework itself and a local desktop app called MemHub, all under the Apache 2.0 license. The release, published on October 16, targets a bottleneck that has quietly become expensive as agents spread: extracting, updating and reusing long-term memory by calling a frontier LLM on every write.

The model layer is Hy-Memory-30B-A3B, a mixture-of-experts model with roughly 30 billion total parameters and about 3 billion active ones, trained specifically for memory extraction and conflict handling rather than conversation. Tencent says it was trained in two stages: on-policy distillation, in which the student model learns memory-write protocols on states it generates itself, and attribution-based reinforcement learning that uses per-block probes to measure how much each stored memory actually contributed to downstream answers.

The framework handles three memory types. Chat memory now groups scattered facts into events with time, place and participants, merging duplicates and resolving conflicts before anything reaches the model. Working memory and code memory are new in 2.0: working memory separates disposable session facts from reusable lessons and gates writes behind a value threshold, while code memory stores file, function and branch coordinates plus diff version chains instead of entire files, keeping the memory base from ballooning.

On benchmarks, the numbers are Tencent's own and should be read accordingly. On a fixed 20 percent subsample of tests spanning Mem0, Hy-Mem, MemOS and EverMemOS, the company reports a macro average of 55.49 for its memory model, above Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64, and it holds the lead on the held-out EverMemOS set at 56.89 versus 54.04. Working memory averages 55.31 percent on MemoryArena travel-planning and shopping tasks, code memory reaches 64.2 percent question-answering accuracy on CodeMemoryBench, and chat memory hits 72.6 percent on LoCoMo while cutting tokens per question from 1,159 to 703, a 39.3 percent reduction.

The third layer, MemHub, is a local-first command-line tool and Mac client that imports past sessions from authorized AI tools into one local memory store and keeps syncing new ones. Tencent says it currently connects ten tools, including WorkBuddy, Cursor, Codex, Qwen Code, OpenCode and DeepSeek Harness, so a preference learned in one coding assistant carries into another. Conflicting memories are not deleted: they get superseded pointers and lower retrieval weight, with old entries kept for traceability.

Tencent says the stack is already used internally across finance, news, knowledge management, office agents, social entertainment and games. The strategic point is the repositioning: memory is being pulled out of the flagship model and turned into cheap, pluggable infrastructure, with a 3B-active specialist doing the high-frequency writing that would otherwise burn frontier-model tokens. The obvious caveats apply: the headline comparison against Opus 5 comes from the vendor on a subsample it chose, MemHub ships for Mac only with no Windows client yet, and real-world behavior over months of accumulated memory remains unproven until independent teams replicate the results.

Comments (0)

Log in to join the discussion

Log In

No comments yet