mem0 external memory system: three-tier scopes and automatic fact extraction

On October 3, 2026, an article on the developer platform Juejin introduced mem0, an external memory system. According to that article, mem0 rests on two core capabilities: layered memory scopes and automatic fact extraction.

The memory layer is split into three scopes, named user_id, agent_id and run_id. The article maps them to the user level, the Agent level and the session level respectively: user_id covers memory belonging to a user, agent_id covers memory belonging to an Agent, and run_id covers memory belonging to a single session. This three-way split is the main description the article offers of how the system is structured.

Automatic fact extraction works by handing a stretch of conversation to an LLM, which then pulls out the key facts on its own. As the article puts it, you drop a conversation in and the LLM extracts the key points, meaning the selection of facts is left to the model rather than to manual cleanup of the dialogue beforehand.

The coverage stays at the level of introducing the concept and its core capabilities. It does not give a release or launch date, a version number, performance figures, or details of how the system is actually used, and it makes no comparison with other memory approaches. So the story currently stops here: mem0 is presented to readers as a system defined by three memory scopes plus automatic fact extraction, with no further verifiable developments beyond that.