The gist

TencentDB Agent Memory is an open-source memory system for AI agents, developed by Tencent. It addresses agents forgetting context in long-running tasks by providing a layered memory architecture. This system helps agents learn workflows, retain task context, and reuse past experience without simply accumulating a massive history or using lossy summarization. It structures memory to improve reasoning and reduce human repetition.

What it does

Condenses verbose tool logs into compact Mermaid symbol graphs to manage short-term memory.
Distills conversations into a layered semantic pyramid of facts, scenes, and personas for long-term memory.
Integrates with agent frameworks like OpenClaw and Hermes as a plugin or adapter.
Provides full traceability from high-level abstractions back to raw logs for debugging.
Uses a local SQLite backend with hybrid retrieval for out-of-the-box functionality.
Generates human-readable memory artifacts (Markdown, Mermaid) for white-box inspection.

How it works

TencentDB Agent Memory operates as a plugin for agent frameworks or a self-hosted Docker container. It automatically captures agent conversations and tool outputs, processing them into a layered structure. For short-term tasks, it creates symbolic Mermaid graphs; for long-term memory, it builds a semantic pyramid. It runs with a local SQLite backend by default and is configured via JSON files. The project is open-source under the MIT license.

Best for

This tool is ideal for developers building long-horizon AI agents who require a persistent, auditable memory system to improve performance and reduce context window costs across extended sessions.

Watch out for

This is a library for developers, not a standalone application. Full functionality requires integration with a supported agent framework like OpenClaw or Hermes, and may involve applying patches to the host framework.