Hy-Memory 2.0
An open-source persistent memory solution from Tencent Hunyuan that uses a three-layer memory architecture (Working Memory, Code Memory and Chat Memory) to manage AI agent memory. It is released under the Apache 2.0 license.
Timeline (last 90 days)
Tencent Hunyuan open-sourced the persistent memory solution Hy-Memory 2.0 under the Apache 2.0 license on GitHub and Hugging Face.
Hy-Memory 2.0 consists of three complementary components: the client library and SDK hy-memory (stable version 2.0.2), the local desktop and command-line management application MemHub, and the neural network model Hy-Memory-30B-A3B.
Hy-Memory 2.0 achieves 72.6% accuracy on the LoCoMo continuous memory evaluation.
Hy-Memory 2.0 reduces the average tokens consumed per query from 1159 to 703, cutting context overhead by 39.3%.
In the LoCoMo evaluation, Hy-Memory 2.0's 72.6% accuracy is higher than GPT-6 Astra's 71.2%, Claude Fable 5.1's 70.8% and DeepSeek 4.1's 68.4%.
On the MemoryArena goal persistence evaluation, the Working Memory layer scores 55.31%, 8.4 percentage points higher than a pure summarization compression pipeline.
On CodeMemoryBench, the Code Memory layer achieves 64.2% code context recall accuracy and a 35.6% direct resolution rate for SWE-level issues.
Hy-Memory 2.0 writes memories to a local SQLite relational database and a hierarchical vector index, and the official SDK provides an asynchronous interface that can read and write memories with millisecond-level latency on transaction-protected relational storage.
In local SQLite, Hy-Memory 2.0 can return already-cleaned facts within 4 milliseconds.
All Facts (9)
In local SQLite, Hy-Memory 2.0 can return already-cleaned facts within 4 milliseconds.
50%UnverifiedHy-Memory 2.0 consists of three complementary components: the client library and SDK hy-memory (stable version 2.0.2), the local desktop and command-line management application MemHub, and the neural network model Hy-Memory-30B-A3B.
50%UnverifiedHy-Memory 2.0 achieves 72.6% accuracy on the LoCoMo continuous memory evaluation.
50%UnverifiedHy-Memory 2.0 reduces the average tokens consumed per query from 1159 to 703, cutting context overhead by 39.3%.
50%UnverifiedIn the LoCoMo evaluation, Hy-Memory 2.0's 72.6% accuracy is higher than GPT-6 Astra's 71.2%, Claude Fable 5.1's 70.8% and DeepSeek 4.1's 68.4%.
50%UnverifiedOn the MemoryArena goal persistence evaluation, the Working Memory layer scores 55.31%, 8.4 percentage points higher than a pure summarization compression pipeline.
50%UnverifiedOn CodeMemoryBench, the Code Memory layer achieves 64.2% code context recall accuracy and a 35.6% direct resolution rate for SWE-level issues.
50%UnverifiedHy-Memory 2.0 writes memories to a local SQLite relational database and a hierarchical vector index, and the official SDK provides an asynchronous interface that can read and write memories with millisecond-level latency on transaction-protected relational storage.
50%UnverifiedTencent Hunyuan open-sourced the persistent memory solution Hy-Memory 2.0 under the Apache 2.0 license on GitHub and Hugging Face.
50%