--- title: "Tencent Mixyuan releases Hy-Memory" type: "News" locale: "en" url: "https://longbridge.com/en/news/288032606.md" description: "On May 28th, Tencent officially launched Hy-Memory, a memory plugin designed for long-term collaborative Agents, aimed at solving the problem of fragmented memory. Hy-Memory adopts a 6-layer memory framework and dual-system design, significantly improving the density and update speed of memory information, with a 35% reduction in token consumption when processing ultra-long contexts. This plugin better meets the long-term memory needs of Agents and enhances user experience" datetime: "2026-05-29T08:13:03.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/288032606.md) - [en](https://longbridge.com/en/news/288032606.md) - [zh-HK](https://longbridge.com/zh-HK/news/288032606.md) --- # Tencent Mixyuan releases Hy-Memory According to Zhitong Finance APP, on May 28th, Tencent Hunyuan officially launched Hy-Memory. It is described as a memory plugin specifically designed for long-term collaborative agents like Openclaw, which can truly become the agent's "second brain." The official statement claims that Hy-Memory utilizes a 6-layer memory framework × System1/System2 dual system × three-layer evolutionary chain, allowing the agent to truly "remember, remember correctly, remember lightly, and understand you better" during long-term use. In authoritative public test sets, Hy-Memory outperformed existing mainstream memory frameworks, solving the problem of fragmented memory, reducing the number of memories by over 70%, and increasing the density of each memory information by over 45%. In terms of handling ultra-long contexts, the token consumption of Hy-Memory is reduced by 35%, and the memory update speed is 20% faster. ## The requirements for long-term memory in Agent tasks far exceed Chat Users who deeply engage with Agent and Openclaw often describe a common "three-week trajectory." First week: Honeymoon period, where users share everything they are currently busy with, such as the background of a project, recent decisions and trade-offs, and future directions, all at once with Openclaw. It can answer questions, look up information, help with planning, write code, and draft documents. "This thing is really useful" — this is the genuine feeling of most people in the first week. Second week: Beginning to feel uneasy. People notice that every day when they open Openclaw, they have to spend 3 to 5 minutes reminding it what they are doing. It seems to forget the judgments discussed a few days ago. When you say to it, "according to the previous plan," it will ask, "which plan?" When you say, "the option we excluded," it cannot remember what was excluded or why it was excluded. It doesn't completely forget. Openclaw's default memory mechanism can remember some recent conversation transcripts, but for deeper judgments that span days or sessions, it drifts away. Third week: Actively downgrading usage, starting to subconsciously shorten the depth of discussions with it. No longer asking, "Should I pursue this direction?" because they know it will forget by tomorrow. It turns into only asking the most specific, immediate small questions: searching for information, checking grammar, revising a paragraph. The most common outcome is that Openclaw gradually degrades in the eyes of users from "a partner that can think with you" to "a query tool." This is not Openclaw's fault — its core capabilities have always been strong. The problem lies in the fact that long-term collaborative tasks require long-term memory far beyond that of ordinary chat. The original intention of the Hy-Memory plugin for Openclaw is to erase the last two weeks of this three-week trajectory — allowing users to use Openclaw in the same way in the first week and the third month, and to understand it better the more they use it **1\. The challenge of long-term tasks on memory is harder than imagined** Long-term tasks are completely different from regular chats. Whether you are using Openclaw to research a complex decision, write a book, plan a major trip, or follow up on a cross-quarter project, a single conversation may involve 30, 50, or even 100 rounds, continuously checking information, adjusting tools, reading documents, proposing solutions, reverting, modifying, and proceeding. A session can last several hours, and tasks may span several weeks. The memory system must handle at least four things: ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/101a3475d2ad535b15a77b8788e9fa62.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) If any link breaks, the experience breaks. **2\. Three standards a qualified Openclaw memory plugin must meet** When designing Hy-Memory, we established three hard standards around the question, "What kind of memory plugin is worthy of the long-term collaboration scenario of Openclaw?" Standard One: No loss of history. All "why this, why not that" discussions between the user and the agent—these judgments and causal relationships cannot be lost. Otherwise, the next time it recommends a solution, it might directly suggest something the user has already rejected, wasting time. Standard Two: Must be able to evolve. Humans are not static beings. Technical preferences, lifestyle habits, and long-term goals change over time, and views on certain matters will also be adjusted repeatedly; all of these must be remembered. It should not be a covering approach of "only remember the latest," nor a piling approach of "keep everything," but rather have a clear evolutionary trajectory. Standard Three: Must be fast enough in the main workflow, with cognitive iteration and evolution. The memory plugin must not slow down the user's current typing, tool invocation, or result waiting speed. Each search must be in the millisecond range, and recall should not become a response bottleneck. At the same time, it must have deep semantic and factual understanding. Next, we will clarify the three core layers of Hy-Memory that meet these three standards. ## The Three Core Layers of Hy-Memory **Layer One: Six-layer memory framework—finding the right place for each type of memory** ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/61b16c16a6922218143cd957943225d3.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) The first thing Hy-Memory does is not to stuff all memories into the same table. Think about it—when a user talks to Openclaw about preparing to study abroad, this statement actually mixes several types of information: "I am preparing to study abroad"—this is a fact. "I prefer the slow-paced projects of Northern Europe"—this is a profile "I will first list the pros and cons before making a major decision" - this is a mental model. "I will probably ask how to find a professor for a recommendation letter next week" - this is a forward-looking intention. These pieces of information are completely different in form, but traditional memory systems mix them all into a pile of vectors, retrieving them all at once. Hy-Memory divides memory into 6 layers, each with a specific responsibility. Below is a set of mixed examples from multiple domains to help you feel what the Agent "remembers": ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/f0705a52656f43da08d5ec8a610e9b77.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) When a user asks Openclaw "What habits do I have when making major decisions," it should prioritize looking at L5 mental models, rather than dumping the last 100 rounds of dialogue verbatim. When the user asks "Where do I live now," a single fact from L2 is enough. What to ask, which layer to go to, and what kind of retrieval weight to use - layering makes all of this possible. The entire prompt becomes cleaner. The model's attention is no longer diluted by irrelevant original text. **Second Layer: Dual System Design, Ensuring Speed and Cognitive Iteration** Hy-Memory directly replicates the System1/System2 mechanism from cognitive science into the memory processing of the Agent - it's like giving Openclaw a "brain" that aligns with cognitive science. Hy-Memory breaks down memory processing into two systems. System1 (Day Shift): Processes the written memory in real-time the moment the user hits enter. Responsible for writing raw traces, extracting facts, updating profiles, and compressing conversation summaries - that is, L1–L4. System2 (Night Shift): Runs in the background from seconds to minutes. Responsible for extracting mental models, building knowledge networks, and predicting intentions - that is, L5–L6. Why break it down? Because deep cognition is slow. Extracting your "decision mental model" and building your "knowledge network" - completing one round may take LLM 5 to 20 seconds. If you had to wait 20 seconds for a response every time you called Openclaw, no one would use it. But what you want is not just "fast" - you want the Agent to understand you better over time. The split in Hy-Memory turns these two conflicting tasks into two independent channels: 1. After sending a message, System1 has already written the "immediately available memory" - the next dialogue can use it right away. 2. System2 slowly deepens cognition in the background - solidifying your two weeks of dialogue into "your decision mental model." The result has a direct impact on dialogue: it can immediately remember the information you just shared; meanwhile, its understanding of you continues to deepen in the background **Layer Three: Evolution Chain - Memory Can Be Rewritten but Does Not Lose Causality** The third layer is the real killer feature of Hy-Memory, and it is also what we believe is the easiest pitfall for users to fall into when collaborating with the Agent long-term, as well as the most worthwhile aspect to get right. A typical long-term scenario: Suppose a user has been discussing their fitness plan with Openclaw for over half a year. During the past six months, their training approach has undergone four significant attitude shifts: Last Spring - Started running for aerobic training, saw good results, improved weight and mental state, full of confidence in running. Last Summer - Joined HIIT training, intensity was too high, knee problems arose, forced to stop training for two weeks. Last Autumn - Shifted to pure strength training, only lifting heavy weights, gave up cardio, ended up gasping after running two kilometers, feeling frustrated. Last Month - Formulated a mixed plan of "strength + jogging + yoga," and finally stabilized their physical condition. Today, they open Openclaw and ask, "I want to add a new training method next month, what do you suggest?" **Three memory systems will provide three different answers.** **Answer 1: Coverage Approach (Only Remember the Latest)** Many default memory systems operate on "new covers old." The database only retains "the user is currently doing well with mixed training." The Agent's response: "Your current combination of strength + jogging + yoga is great; you can try adding CrossFit for efficiency." Not incorrect, but very shallow. It does not understand why the user arrived at mixed training—missing the complete path of "knee injury from HIIT" and "losing cardio from pure strength." Directly recommending CrossFit could likely lead to another knee injury for the user. **Answer 2: Accumulation Approach (Retain Everything)** Some systems retain "everything, recalling based on vector similarity during searches." All four facts are in the database, but when searching for "new training methods," the most recent and similar ones are recalled first—most likely "running started well" and "mixed training is stable." The intermediate "knee injury from HIIT" and "losing cardio from pure strength" are slightly further away in semantic space and may not be recalled at all. The Agent's response: "You had good results with running, and mixed training is stable; you can try adding HIIT to improve cardio efficiency." This is a misstep. It fails to recognize that the user has already been injured from HIIT—recommending HIIT again means leading the user down the same old path, risking another knee injury. **Answer 3: Hy-Memory Evolution Chain** Hy-Memory, when writing the second memory, will use a supersedes pointer to point to the first; the third points to the second; the fourth points to the third. The four memories are linked together by pointers. D (mixed training, chain head) ↓ supersedes C (pure strength, lost cardio) ↓ supersedes B (HIIT, knee injury) ↓ supersedes A (running started) User asks, "What new training method should I choose?" At the time, when the search hits the chain head D, the entire chain automatically expands—A, B, C, and D return together to the Agent. ## Hy-Memory: Exceeding Similar Frameworks Hy-Memory has been tested on two authoritative long-term memory evaluations: LongMemEval and PersonaMem. 1. LongMemEval: Scored 85.2, far exceeding other frameworks. LongMemEval is recognized as a very difficult long-term memory evaluation, covering 500 questions across 6 ability dimensions. ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/41cee9c84c4a53c3991614c81eb58314.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) Hy-Memory not only outperformed all similar systems but also achieved the highest scores in 4 out of the 6 dimensions—especially leading in the three areas that test "evolutionary ability": preference (+21.11pp), temporal reasoning (+9.63pp), and knowledge updating (+21.37pp) compared to similar products. 1. PersonaMem: Outperformed all similar products, PersonaMem is a real long-term dialogue evaluation with over 6000 messages / 589 questions, closer to the real scenario of "users using it for several months." ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/2a899e6674d9de7d0de84a3cb488f409.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg)![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/c311b44800a163cc238df69e5a97fd7c.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) 1. Performance: Higher memory density with 8 times the writing speed. ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/f9827c2d867b369bab94285c52218134.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) Openclaw users can feel that: Fast writing: On par with mem0, 8 times faster than Graphiti, without bottlenecking Openclaw's main dialogue link. Less storage: The number of memories is only 1/3 of mem0 and 1/4–1/5 of Graphiti—local embedded storage, no need for external services, low memory usage. High density: Each memory has 3–4 times the information density of mem0, making recalled memories more useful, with prompts not polluted by noise The last point is particularly important for long-term tasks: when a session runs to 80 rounds of dialogue, every recalled memory must be high-density; otherwise, the prompt will be overwhelmed by noise, diluting the model's attention. The high density of Hy-Memory means that under the same token budget, Openclaw can see 3-4 times more effective information than other memory systems. Quickly set up Hy-Memory: from decision to usability, it takes about 5 minutes. 1. Install with a single command npm config set registry https://mirrors.tencent.com/npm/ && openclaw plugins install @tencent/hy-agent-memory --dangerously-force-unsafe-install --force && openclaw hy-memory init Under the default configuration, Hy-Memory uses Chroma as the local embedded vector library, and data is automatically persisted to the local directory. There is no need to install Qdrant first, no need to start any external services, no need to deploy Docker; as long as the API Key for LLM / Embedding is configured, it can be used directly. You can verify the installation status and whether the LLM and Embedding services are connected with the following command: openclaw hy-memory status \[plugins\] openclaw-hy-memory: registered (user: tom001, server: http://127.0.0.1:19527, autoRecall: true, autoCapture: true) OpenClaw 2026.5.26 (10ad3aa) — Built by lobsters, for humans. Don't question the hierarchy. HY Memory Server: http://127.0.0.1:19527 Status: ✓ healthy User ID: tom001 VDB: ok \[chroma\] (collection: agent\_memories\_1024, points: 0) Embed: ok (dims: 1024) LLM: ok SDK Version: 1.2.5 1. Lite / Pro / Ultra: not necessarily "fully open" Hy-Memory splits the same kernel into three levels, enabling them as needed: ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/3ac781ab071d8f939a5dc889ea9bf008.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) Recommendation: For first-time installation, go directly for Pro. Enable MemAgent but do not start the System2 background. The development machine can generally handle it. For real high-frequency use, you can upgrade to Ultra, turn on System2, and both L5 / L6 will be settled. All three levels share the same SDK, and upgrading only requires changing one switch, no need for re-integration. ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/bbde7bc5bf6250179b1b4a7fa7128152.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) If you are an existing Openclaw user: install Hy-Memory, and you will feel that this Agent is possibly the first AI that truly recognizes you ### Related Stocks - [00700.HK](https://longbridge.com/en/quote/00700.HK.md) - [TCEHY.US](https://longbridge.com/en/quote/TCEHY.US.md) - [TCTZF.US](https://longbridge.com/en/quote/TCTZF.US.md) - [80700.HK](https://longbridge.com/en/quote/80700.HK.md) - [HTCD.SG](https://longbridge.com/en/quote/HTCD.SG.md) ## Related News & Research - [Small biotech firms quicker to 'latch onto' AI than big ones, says Tencent Healthcare President](https://longbridge.com/en/news/287909875.md) - [TikTok Owner ByteDance Is Reportedly Offering AI Staff A Sweet Deal As Talent War Heats Up](https://longbridge.com/en/news/287641148.md) - [Tencent links PayPal to WeChat Pay network, enabling US users to spend across China](https://longbridge.com/en/news/287766313.md) - [XLM surged 28% on DTCC news: can the rally continue?](https://longbridge.com/en/news/288017834.md) - [Can AI help Hongkongers quit smoking? Health authorities are giving it a go](https://longbridge.com/en/news/287754035.md)