Tencent Open-Sources Agent Memory Technology, Token Consumption Reduced by Up to 61%
Tencent Cloud has open-sourced TencentDB Agent Memory, a memory compression solution for long-duration agent tasks that reduces token usage by up to 61% and improves task success rates by up to 51%. It features a Mermaid task canvas and context offloading to external storage.
On May 14, Tencent Cloud announced the global open-sourcing of TencentDB Agent Memory, a memory compression technology designed for long-duration agent tasks. The solution provides both long-term and short-term memory compression capabilities, enabling seamless deployment with mainstream agent frameworks such as OpenClaw and Hermes. In multi-task continuous session experiments, the technology achieved up to a 61% reduction in token consumption and a 51% increase in task success rates.
As agents increasingly handle complex, multi-step tasks in code development, web search, and research analysis, the volume of tool-generated data—including web content, code logs, and intermediate results—quickly fills the context window, driving up token costs. Common approaches like expanding context windows or compressing historical content often still lead to context bloat and task state confusion. TencentDB Agent Memory proposes a novel memory compression strategy: offloading complete information to external storage while preserving key states and structural relationships during task execution.
One of its core features is the "Mermaid task canvas," which organizes the agent's task process into a structured task graph that retains task states, step summaries, and execution relationships. By retaining only this lightweight task canvas in the context, the agent can quickly assess progress and dependencies between steps. Another key capability is "Context Offloading," which moves raw data such as web content, logs, and intermediate results to an external file system after tool calls, keeping only summaries and index information in the context. Data can be restored on demand when needed.
Test results from Tencent Cloud show that with this memory scheme, token consumption in web search scenarios drops by up to 61% and success rates improve by 52%. In code repair scenarios, token consumption decreases by up to 33% and completion rates increase by 10%. In long, complex document scenarios, token consumption is reduced by up to 31% and accuracy improves by 8%.
Additionally, the project includes a long-term personalized memory module that automatically extracts user preferences and historical background across agent conversations. In PersonaMem evaluations, this module boosted the agent's accuracy in understanding user profiles from 48% to 76%. The module was previously offered as a free service.
TencentDB Agent Memory is now open source on GitHub, compatible with OpenClaw, Hermes, and other major agent frameworks. Developers can install it with a single command without needing to configure a separate database or external service. Task records, memory content, and other information are saved as plain files, making them easy to view and debug.
This year, Tencent has been accelerating its agent technology ecosystem. Last month, Tencent Cloud open-sourced Cube, an agent execution foundation, which garnered over 5,000 GitHub stars within two weeks.