Tencent Open-Sources Agent Memory Technology, Cutting Token Use by Up to 61%

Tencent open-sources Agent Memory Solution, Reducing Token Consumption by Up to 61%

May 14 – Tencent has officially open-sourced TencentDB Agent Memory, a new mEMOry solution designed for long-task AI agent scenarios, according to a post on Tencent Cloud’s official WeChat account. The solution provides both long-term and short-term memory compression capabilities and supports one-CLIck deployment on mAInstream Agent Frameworks such as openclaw and Hermes.

In multi-session continuous task experiments, TencentDB Agent Memory reduced token consumption by up to 61% and improved task success rates by as much as 51% compared to baseline methods.

As AI Agents are increasingly APPlied to code development, Web Search, research analysis, and other long-horizon tasks, intermediate results—such as web pages, code logs, and tool ouTPUts—quickly fill up context Windows, driving up token costs. Common industry approaches, such as enlarging context windows or summarizing historical content, often lead to context bloat and task state confusion in long-running scenarios.

TencentDB Agent Memory introduces a novel memory compression architecture: full Information is offloaded to external stoRAGe, while key task states and structural relationships are preserved.

Two core capabilities:

1. Mermaid task canvas – The agent’s task process is organized into a structured task graph that retains task status, step summaries, and execution dependencies. Even with only a Lightweight task canvas in context, the agent can quickly underStand progress and inter-step relationships.

2. context offloading – After tool calls are completed, raw information (web content, logs, intermediate results) is no longer kept in the context window. Instead, it is stored in an external file system, with only summaries and index references kept in context. Full data can be restored on demand.

Test results released by Tencent Cloud show:

• Web search scenarios: token consumption reduced by up to 61%, task success rate increased by 52% (relative).

• Code repair scenarios: token consumption reduced by up to 33%, task completion rate increased by 10%.

• Long document scenarios: token consumption reduced by up to 31%, accuracy increased by 8%.

The project also includes a long-term personalized memory module that automatically extrACTs user preferences and historical context across conversations. On the PersonaMem benchmark, the module improved agent accuracy in user profile understanding from 48% to 76%. This module was previously released as a free service.

Agent Memory is now open-source on GitHub and supports mainstream agent frameworks including OpenClaw and Hermes. Developers can install it with a single command—no additional database or external service configuration required. Agent task records and memory content are saved as plain files for easy viewing and debugging.

Tencent has been accelerating its AI Agent Technology ecosystem this year. Last month, Tencent Cloud open-sourced its agent execution base, Cube, which surpassed 5,000 GitHub stars in two weeks.