TencentDB Agent Memory

TencentDB Agent Memory is an open-source memory layer for AI agents that rejects flat vector dumps in favor of a four-tier progressive pipeline (L0 Conversation, L1 Atom, L2 Scenario, L3 Persona). Short-term tool logs are offloaded to external files and condensed into a lightweight Mermaid canvas the agent can reason over, while long-term knowledge is distilled into structured personas and scenes with drill-down paths back to raw evidence. Hybrid retrieval combines BM25 and vector search over a local SQLite backend with zero external API dependencies.

It is aimed at developers building long-horizon autonomous agents where context bloat and irreversible summarization are real problems. Tencent publishes benchmark numbers on WideSearch, SWE-bench, AA-LCR, and PersonaMem showing token reductions in the 30-60% range and substantial accuracy gains when paired with the OpenClaw agent framework. The project is MIT-licensed, written primarily in TypeScript (Node 22+), and ships as an npm package with HTTP endpoints for capture, search, and recall.

Integrations include a first-party OpenClaw plugin and a Hermes Gateway adapter. Because everything runs locally, it is suitable for teams that cannot send agent traces to managed memory services, though it is not a hosted product and requires engineering effort to wire into an existing agent stack.