en.Wedoany.com Reported - On May 14, Tencent Cloud officially open-sourced TencentDB Agent Memory, providing short-term memory compression and long-term personalized memory capabilities for Agent long-task scenarios. Developed by the Tencent Cloud Database Team, the project is open-sourced under the MIT license, ready to use out of the box, and has already been adapted to mainstream Agent frameworks such as OpenClaw and Hermes.
The core technical pathway of TencentDB Agent Memory consists of two mechanisms: "Context Offloading" and "Mermaid Task Canvas." Context Offloading writes the complete results to an external file after each tool call, retaining only a one-line summary and index path in the context, so the original information no longer occupies the context window long-term. The Mermaid Task Canvas uses flowcharts to organize the task execution process into a navigable, structured map, allowing the Agent to know only the organizational location and expansion path of key information without needing to remember everything. Ablation experiments show that Token savings are approximately 15% with offloading alone, increasing to 31% to 33% when combined with the Mermaid Canvas.
In ultra-long session evaluations, after integrating TencentDB Agent Memory as an OpenClaw plugin, the success rate on the WideSearch benchmark increased from 33% to 50%, a relative improvement of 51.52%, while Token consumption dropped from 221.31M to 85.64M, a reduction of 61.38%. The SWE-bench success rate rose from 58.4% to 64.2%, with Token consumption reduced by 33.09%. The AA-LCR success rate improved from 44% to 47.5%, with Token consumption reduced by 30.98%. In the long-term memory dimension, PersonaMem evaluation accuracy increased from 48% to 76%.
TencentDB Agent Memory builds a four-layer progressive memory architecture. The L0 Raw Dialogue Layer fully retains every interaction round. The L1 Atomic Memory Layer automatically extracts facts, preferences, constraints, and stage conclusions. The L2 Scene Summarization Layer automatically aggregates by task. The L3 User Profile Layer continuously distills stable long-term profiles. The layers are connected through extraction-aggregation-distillation pipelines, and any layer can be independently upgraded or replaced.
The product's long-term memory feature was launched for free use in April 2026. That same month, Tencent Cloud Storage Expert Architect Wang Dengyu proposed the Agent Memory Lake concept at the 2026 Artificial Intelligence Infrastructure Summit, benchmarking against data lakes to build a unified memory foundation for agents, spanning the entire process of agent understanding, reasoning, execution, and reflection. On the underlying model side, the Tencent Hunyuan Hy3 preview version was released and open-sourced in late April 2026, achieving significant improvements in complex reasoning, long memory, multi-turn questioning, and Agent capabilities, forming a dual support with TencentDB Agent Memory at the model and engineering levels in the memory dimension.
TencentDB Agent Memory uses local SQLite storage by default with zero external dependencies. Users can complete installation by executing openclaw plugins install@tencentdb-agent-memory/memory-tencentdb via the command line. Tencent Cloud also announced the upcoming launch of the Agent Memory Pro version for multi-user and enterprise-level scenarios, built on the Tencent Cloud Vector Database, supporting data governance capabilities such as backup, rollback, and permission control.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
