Hy-Memory 2.0
Model, framework and app — all open-source
Help AI remember what you said, how things unfolded, where your work stands, and why the code changed. Each layer works on its own, and together they form a complete memory solution — your personal memory steward.
Three open-source layers, each with its own focus, built to work together
Upgraded chat memory, plus new working and coding memory
Hy-Memory is the first memory framework in the industry to introduce working memory and coding memory.
Content across time and conversations is linked and organized: each event is structured around key elements such as time, place and people, and connected to how it later unfolds. When an answer needs several clues, duplicates are merged and contradictions resolved first, so the model receives clean, clear memory.
Remembers a task along multiple dimensions and keeps every update in time order: goals, settled approaches and interim results, plus reusable experience and the conditions under which it applies. Before reusing experience, it checks whether those conditions hold, and applies it only when they do.
For long-horizon tasks, each time work resumes it first identifies what is being worked on, then retrieves the relevant decisions, still-applicable experience and past work history — fetching exactly the memory that helps the task move forward.
Note · The underlying design of 1.0 is described in arXiv 2606.09483, where it is named DCPM.
Minimal storage across all three memory scenarios, #1 performance among open-source frameworks
On 4 benchmarks spanning chat, working and coding scenarios, Hy-Memory ranks #1 among open-source frameworks against Mem0, MemOS, EverOS, OpenViking, RepoMemory and Simple RAG. Just as important, it uses very little storage in all three scenarios.
MemoryArena
On MemoryArena's travel-planning and bundled-shopping tasks, Hy-Memory averages 55.31%, 10.2 percentage points ahead of the runner-up, EverOS.
Note · Average = mean of Group Travel Planner SPS and Bundled Shopping PS.
X-axis: memory storage footprint (Bank tokens ÷ Input tokens); Y-axis: score in that scenario. † Simple RAG stores windowed raw text without LLM extraction, so its storage footprint is not directly comparable with extraction-based memory.
~3B active parameters, a higher average score than Opus5
Trained specifically for memory scenarios, with strong memory updating and management capabilities. Tested across four memory systems — Hy-Memory, Mem0, MemOS and EverMemOS — Hy-Memory-30B-A3B scores 55.49, ahead of Opus5 (54.62), DeepSeek V4.1 Flash (53.64) and Qwen3-30B-A3B (44.37).
| Model | LoCoMo | LongMemEval | PersonaMem v2 | Travel | Shopping | CMB | Avg. |
|---|---|---|---|---|---|---|---|
| Hy-Memory-30B-A3BMemory-specialized | 67.27 | 77.75 | 52.09 | 48.22 | 19.90 | 67.74 | 55.49 |
| Opus5Closed-source | 70.23 | 78.87 | 51.58 | 40.59 | 21.11 | 65.36 | 54.62 |
| DeepSeek V4.1 FlashOpen-source | 66.32 | 77.77 | 49.75 | 41.65 | 20.00 | 66.36 | 53.64 |
| HY3General | 60.59 | 78.75 | 48.88 | 40.32 | 19.30 | 61.95 | 51.63 |
| Qwen3.6-35B-A3BOpen-source | 61.35 | 77.96 | 48.72 | 37.45 | 18.58 | 58.16 | 50.37 |
| HY3-A3B BaseBase | 59.83 | 72.00 | 48.41 | 34.27 | 18.47 | 52.95 | 47.66 |
| Qwen3-30B-A3BOpen-source, same size | 59.35 | 67.75 | 44.69 | 33.16 | 17.64 | 43.64 | 44.37 |
Unit: %. Darker cells indicate higher scores within a column; bold blue marks the column's best. Avg. is the equal-weighted mean of the six metrics; “Average of 4 systems” averages each metric across the four systems.
Note · The six metrics are LoCoMo, LongMemEval, PersonaMem v2, Travel, Shopping and CMB, equally weighted; higher is better. Hy-Memory, Mem0 and MemOS were used in RL training; EverMemOS was unseen in training and is used to test generalization.
All your agents. One Memory.
One personal memory, shared across 10 AI tools. Preferences you explained in Cursor are unknown to Codex; conventions you set in OpenCode must be explained all over again in WorkBuddy. MemHub turns your memory into a long-term asset that no single tool can lock in — so every conversation starts with an AI that already knows you.
Authorize your apps in MemHub to import past sessions from each tool. The built-in Hy-Memory SDK distills them into local memory, so you never start from scratch.
From then on, new sessions in every tool sync in real time into the same memory; a new memory formed in any tool is immediately visible to all the others.
Connect each AI tool in one click as a plugin. Whenever you start a task in a connected tool, MemHub manages and applies the relevant memories as added context.
With MemHub: fresh conversations, lasting rapport.
Deployed across Tencent products, with strong results in dedicated evaluations
Hy-Memory powers Tencent Stocks, iWiki, Tencent News and other products, significantly improving memory recall and answer accuracy so that AI assistants better understand and carry forward users' long-term preferences.
The assistant must remember, across days and sessions, the focus areas, analysis habits and rules users give it — such as “only watch non-ferrous metals and semiconductors this week” or “report the historical P/E percentile first.” When users shift focus, old directions are archived but traceable, and the assistant never pushes abandoned sectors back.
Note · Dedicated memory evaluation for China A-share investment research. Memory model: DeepSeek V4 Flash; baseline: the previous memory solution. Overall score 64.3 → 79.6, a 23.8% relative improvement.
The CTX memory plugin, built on Hy-Memory, lets multiple agents share one long-term memory. 251 questions, same model, same prompts.
The AI remembers the topics users follow and how they like to be addressed, proactively brings up what they care about at the start of a chat, and picks up where the last conversation left off.
Offline evaluation in the Zhisheng scenario: 41 complex conversation cases from 10 users.
How to get started
Three ways in: start from the app, the framework or the model
Sync sessions from multiple AI apps in one click. Distilled through the Hy-Memory pipeline, they come together as one complete personal memory. From then on, every task you start in any app automatically recalls relevant memories to add context to the answer.
GitHub · MemHubUse the SDK to manage cross-session information, task progress and experience.
GitHub · Hy-MemoryTrained specifically for reading and writing memory, it works with Hy-Memory and also plugs into other memory systems such as Mem0 and MemOS.
HuggingFace