{"slug": "hy-memory-benchmark-how-tencent-s-3b-model-beats-claude-opus-5", "title": "Hy-Memory Benchmark: How Tencent's 3B Model Beats Claude Opus 5", "summary": "Tencent Hunyuan released Hy-Memory-30B-A3B, an Apache 2.0-licensed memory-writing model with roughly 3 billion active parameters, which scored 55.49 overall on Tencent's internal memory-writer benchmark, ahead of Claude Opus 5, DeepSeek V4.1, and a size-matched Qwen model that scored 50.37. The benchmark held the retrieval pipeline, prompts, and test cases fixed and swapped only the model handling fact extraction and reconciliation, isolating memory-writing competence from general capability. Tencent trained the model in two stages: a supervised stage in which a frozen Hy-3 teacher model checks the student's memory-writing steps, followed by a reinforcement learning stage that scores the full memory bank using a frozen answer model and a judge; the release includes the Hy-Memory SDK, which supports self-hosted model servers such as vLLM and documented integration with coding agents like OpenClaw.", "body_md": "# Hy-Memory Benchmark: How Tencent's 3B Model Beats Claude Opus 5\n\nTencent's Hy-Memory-30B-A3B scored 55.49 on memory benchmarks, beating Claude Opus 5 and DeepSeek V4.1. Here's how it works.\n\n## What is Hy-Memory and why does it matter?\n\nHy-Memory is an open-source model and SDK from Tencent Hunyuan built for one specific job: deciding what an AI agent should remember, update, or discard. Most [memory frameworks for AI agents](https://goremy.ai/blog/articles/what-is-ai-persistent-memory-a-plain-explainer), including Mem0 and MemOS, rely on a general-purpose LLM to do this extraction and reconciliation work. Tencent instead trained a dedicated model, Hy-Memory-30B-A3B, with roughly 3 billion active parameters, and paired it with its own memory engine and SDK called Hy-Memory. On an internal benchmark comparing memory writers, this smaller specialized model outscored much larger general models including Claude Opus 5 and DeepSeek V4.1.\n\n## TL;DR\n\n- **Tencent released Hy-Memory-30B-A3B** , a mixture-of-experts style model with about 3 billion active parameters, specifically trained to extract and reconcile facts for agent memory systems rather than handle general chat or coding tasks.\n- **The model scored 55.49 overall** on Tencent’s memory-writer benchmark, ahead of Claude Opus 5, DeepSeek V4.1, and a similarly sized Qwen model that scored 50.37.\n- **The benchmark design isolates the writer** , meaning all four models were tested inside the same memory pipeline, with only the model responsible for writing and reconciling memories swapped out.\n- **Training happened in two stages** , a supervised stage where a frozen “Hy-3” teacher model checks the student’s memory-writing steps, followed by a reinforcement learning stage that scores the full memory bank using a frozen answer model and a judge.\n- **The release is Apache 2.0 licensed** , covering both the model weights and the Hy-Memory SDK, which makes it usable in commercial production systems without the licensing friction of some other open releases.\n- **The SDK integrates with existing agent stacks** , including support for pointing it at a self-hosted model server such as vLLM, and documented integration paths for coding agents like OpenClaw.\n\n## Seven tools to build an app. Or just Remy.\n\nEditor, preview, AI agents, deploy — all in one tab. Nothing to install.\n\n## How does Hy-Memory’s benchmark comparison work?\n\nThe benchmark Tencent published tests memory writers under identical conditions. Instead of comparing full chat capability or coding ability, the setup keeps the retrieval pipeline, prompts, and test cases fixed and only changes which model handles the “write to memory” step. That step covers two jobs: extracting the facts worth keeping from a conversation, and reconciling those new facts against whatever is already stored, including deciding whether an old fact should be superseded.\n\nFour models were evaluated this way: Hy-Memory-30B-A3B, Claude Opus 5, DeepSeek V4.1, and a Qwen model matched for size. Hy-Memory scored 55.49 overall, ahead of the Qwen model at 50.37 and ahead of both Claude Opus 5 and DeepSeek V4.1. Because the only variable was the writer model, the result is a direct comparison of memory-writing competence rather than general intelligence. A frontier general-purpose model can still write worse memory entries than a small model trained explicitly for that narrower task.\n\n## Why would a 3B model beat much larger general LLMs?\n\nMemory writing is a narrow, structured task: read new information, compare it against existing stored facts, decide whether something is new, redundant, or contradictory, and then either add, merge, or supersede entries. General LLMs are optimized for broad reasoning, conversation, and task completion, not specifically for this kind of structured bookkeeping over time. A model trained directly on the memory-writing objective can learn the specific failure modes that matter here, such as leaving stale facts in place after an update, duplicating information, or failing to propagate a change to related facts.\n\nTencent’s training approach reflects this. The first stage uses supervised learning where a frozen teacher model, referred to as “Hy-3,” checks each step the student model takes while writing memories. The second stage applies reinforcement learning, scoring the resulting memory bank after each chunk of input using a frozen answer model and a judge, then carrying forward the best-performing memory bank. This means the model is optimized against the actual downstream goal, keeping a memory store accurate over time, rather than against a proxy objective like next-token prediction on generic text.\n\n## How does Hy-Memory handle updating a fact?\n\nA demonstrated test case shows the mechanism clearly. A fact is stored that a person lives in Seattle. A follow-up states that the person moved to Sydney. Rather than appending the new fact alongside the old one, the model issues a “supersede” action: it marks the Seattle memory as outdated, writes out its own reasoning for the change (the person moved and no longer lives in Seattle), and updates related memories, such as job information, to stay consistent with the new state. A subsequent search for “where does this person live” returns Sydney, not Seattle.\n\nThis reconciliation step is the core value proposition. Storing facts is trivial. Correctly deciding when a new fact invalidates an old one, and propagating that change across related entries, is the harder problem that memory frameworks struggle with, especially as a memory store grows and facts drift over time.\n\n## Remy doesn't build the plumbing. It inherits it.\n\nOther agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.\n\nRemy ships with all of it from MindStudio — so every cycle goes into the app you actually want.\n\n## What’s inside the Hy-Memory stack?\n\nHy-Memory ships as two separate but connected pieces:\n\nThe model, Hy-Memory-30B-A3B, is the component that does the actual extraction and reconciliation work. Its mixture-of-experts architecture keeps active parameters low (around 3 billion) while the total parameter count is larger, which keeps inference costs down relative to a dense model of similar capability.\n\nThe SDK, also called Hy-Memory, is the orchestration layer. It handles embedding text into vector representations for storage, manages a lightweight SQLite-backed store, and exposes two primary operations: “add,” which saves new information and triggers the extraction and reconciliation pipeline, and “search,” which retrieves relevant stored memories. The SDK supports a “pro mode” that engages the full LLM-based extraction and reconciliation pipeline, as opposed to simpler modes that may skip reconciliation.\n\nThe SDK is model-agnostic for both the writer and the embedding step. It can point at a self-hosted model served through a framework like vLLM, or at hosted APIs. Embeddings can similarly come from a local model or an API-based embedding service.\n\n## Is Hy-Memory worth using over Mem0 or MemOS?\n\nFor teams already running Mem0, MemOS, or a similar memory framework backed by a general LLM, the appeal of Hy-Memory is cost and consistency rather than raw intelligence. A 3-billion-active-parameter model is cheap to run locally compared to routing every memory write through a frontier model API, and the benchmark results suggest you aren’t sacrificing quality to get that cost reduction, at least on the specific task of memory extraction and reconciliation.\n\nThe Apache 2.0 license is also a practical factor. It allows commercial use without the restrictions that come with some other open-weight releases, which matters for anyone planning to run this in a production agent system rather than a research project.\n\nThe tradeoff is that Hy-Memory is a single-purpose tool. It is not a general chat model and isn’t meant to replace the LLM an agent uses for reasoning or conversation. It plugs into the memory layer specifically. Teams need an existing or new memory architecture (vector store, SQLite backend, retrieval logic) for the model to sit inside, which the Hy-Memory SDK provides out of the box, or which can be adapted into an existing custom pipeline.\n\n## Frequently Asked Questions\n\n### What is Hy-Memory-30B-A3B?\n\nIt’s an open-source language model from Tencent Hunyuan with about 3 billion active parameters, trained specifically to extract facts from conversations and reconcile them against existing stored memories for AI agent systems.\n\n### How does Hy-Memory compare to Claude Opus 5 on memory tasks?\n\nOn Tencent’s published memory-writer benchmark, Hy-Memory scored 55.49 overall, ahead of Claude Opus 5 and DeepSeek V4.1, both of which are much larger general-purpose models not specifically trained for memory writing.\n\n### What license is Hy-Memory released under?\n\nBoth the Hy-Memory-30B-A3B model and the Hy-Memory SDK are released under the Apache 2.0 license, which permits commercial use.\n\n### Can Hy-Memory run on local hardware?\n\n## Other agents start typing. Remy starts asking.\n\nScoping, trade-offs, edge cases — the real work. Before a line of code.\n\nYes. It can be served locally with inference frameworks like vLLM. Running it alongside a usable context length is feasible on a single high-memory GPU, though exact VRAM needs scale with the context length configured.\n\n### Does Hy-Memory replace frameworks like Mem0?\n\nNot entirely. Hy-Memory is a model plus an SDK that handles the memory writing and reconciliation step. It can be used as a drop-in replacement for the general LLM that frameworks like Mem0 or MemOS normally use for that same task, rather than replacing the entire framework.", "url": "https://wpnews.pro/news/hy-memory-benchmark-how-tencent-s-3b-model-beats-claude-opus-5", "canonical_source": "https://www.mindstudio.ai/blog/hy-memory-benchmark-vs-alternatives/", "published_at": "2026-10-23 00:00:00+00:00", "updated_at": "2026-10-23 11:12:58.712831+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-tools", "ai-research"], "entities": ["Tencent", "Tencent Hunyuan", "Hy-Memory-30B-A3B", "Hy-Memory", "Claude Opus 5", "DeepSeek V4.1", "Qwen", "OpenClaw"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/hy-memory-benchmark-how-tencent-s-3b-model-beats-claude-opus-5", "markdown": "https://wpnews.pro/news/hy-memory-benchmark-how-tencent-s-3b-model-beats-claude-opus-5.md", "text": "https://wpnews.pro/news/hy-memory-benchmark-how-tencent-s-3b-model-beats-claude-opus-5.txt", "jsonld": "https://wpnews.pro/news/hy-memory-benchmark-how-tencent-s-3b-model-beats-claude-opus-5.jsonld"}}