Tencent open-sources Agent memory system, OpenClaw saves up to 61% of tokens

AIMPACT News: On May 14 (UTC+8), according to Dongcha Beating monitoring, the Tencent Cloud Database team spent 6 months specifically tackling the long-conversation forgetfulness issue and has recently officially open-sourced TencentDB Agent Memory. This is a local-first memory engine for AI Agents. By default, it uses SQLite + sqlite-vec as the local backend, and it can be installed as an OpenClaw plugin; it also supports integration with Hermes Gateway.

Its core is not to directly put historical conversations into a vector database, but to split memories into two structures. For long-term memory, it is layered as L0 raw conversations, L1 atomic facts, L2 scene chunks, and L3 user profiles. For short-term task memory, it externalizes lengthy tool logs into refs files, writes step summaries into jsonl, and uses a Mermaid canvas to preserve the task structure and node indices. In complex workflows with more than 30 steps, the Agent usually only reads the lightweight Mermaid structure diagram, and when it needs to verify details, it goes back to the original logs via node_id.

According to official benchmarks, after integrating OpenClaw, the token consumption for the WideSearch task decreased from 221.31M to 85.64M (a 61.38% reduction), while the success rate increased by 51.52% relative to before. In the long-term memory evaluation PersonaMem, accuracy improved from 48% to 76%. The value of this design is that it does not swallow historical details with a one-time summary; instead, it preserves the complete path from high-level profiles and the task canvas all the way back to the underlying original text.

(Source: BlockBeats)