Tencent Hunyuan has launched and open-sourced HyMemory 2.0, a framework designed to give AI agents working, code and chat memory. It also offers a local memory app that can connect to ten AI tools, making this a practical attempt to help agents carry useful context between tasks rather than start from a blank page each time.
Tencent AI Watch analysis
What happened
The announcement, reported by KuCoin citing IT Home, describes three parts: a memory model, a framework developers can integrate into agents through an SDK, and MemHub, a local memory-management app. The components can be used separately or together. Tencent says HyMemory is already used across areas including finance, news, knowledge management and office agents.
HyMemory 2.0 adds working and code memory alongside chat memory. For longer tasks, it is designed to retain objectives, decisions and progress, then retrieve relevant details when work resumes. MemHub can import past sessions from connected apps, keep new sessions synchronised and supply relevant memories as context. The report names ten supported tools, including Cursor, Codex, Qwen Code, Hermes, DeepSeek Harness and OpenCode. The SDK and MemHub are on GitHub; model weights are on Hugging Face.
Key findings
- Working memory:
Tencent reports an average 55.31% on two MemoryArena tasks, 10.2 percentage points ahead of the next framework named in its comparison. - Code memory:
The report gives scores of 64.2% for code question-answering accuracy and 35.6% for task resolution on CodeMemoryBench. - Chat memory:
Tencent reports 72.6% accuracy on LoCoMo and 703 average tokens per question, down from 1,159. - The model:
The report says Hy-Memory-30A3B scored 55.49% across four benchmarks, ahead of Claude Opus 5 at 54.62% in the stated comparison.
Why it matters
Agent memory is becoming less about saving a transcript and more about finding the right detail when a task resumes. If the framework works as described, developers can add memory for ongoing work and code, while MemHub gives users a way to connect context across tools. That is a useful answer to the agent habit of forgetting everything just when the project gets interesting.
The performance figures are reported claims, not an independent comparison. They are tied to named benchmarks and specific test setups, so they are a starting point for evaluation, not a universal ranking of which memory system will work best in a real application.
Our read
The appealing part is the combination: an SDK for developers, a local app for connecting tools, and separate memory types for chat, work and code. That makes HyMemory more than a leaderboard entry. The test for MemHub will be whether its integrations and memory controls are clear enough to make cross-tool continuity useful without turning a user’s history into an indiscriminate pile of context.
What to watch
- Whether developers publish reproducible results for the reported benchmark comparisons.
- How MemHub handles permissions, deletion and the boundaries between connected apps.
- Whether the ten integrations work smoothly in sustained, real-world tasks.
Discussion spark: Would you connect several AI tools to one shared memory system, or is keeping each tool’s context separate worth the extra repetition?
Sources and evidence
- Tencent Hunyuan Launches Open-Source HyMemory 2.0 Framework, Claims Superior Performance Over Opus 5 – KuCoin (16 October 2026, 12:47 UTC)
not affiliated with or endorsed by Tencent AI