Alibaba Cloud's Project Hub published a write-up on 3 November of Qwen Memory Agent, an MIT-licensed open-source system from its Global AI Hackathon series, deployed on ECS with Qwen Cloud. Memory is exposed to the model as function-call tools, so the agent decides when to remember, recall or forget. A persistent engine stores typed records plus vectors, and retires superseded facts using exact subject and type matching combined with semantic similarity rather than letting duplicates accumulate. Retrieval ranks candidates by similarity, recency, salience and type, then packs them into a configurable token budget; the same engine is served over HTTP and as eight MCP tools, with JSON and Markdown export and import. A separate "dreaming" pass proposes memory consolidation, but changes require human approval. The published results split cleanly: on a six-persona, 24-query synthetic offline fixture, context recall was 1.000 with staleness 0.000 across 8- to 64-token budgets, while a separate ten-scenario live evaluation reached 0.60 task success, where decision outcome, memory-store state and recall-before-decision all had to pass.
Our takeThe two numbers measure different things, and the write-up says so: retrieving the right fact does not mean the agent will use it. Perfect recall on a 24-query fixture is a property of the fixture; 0.60 across ten live scenarios is a property of the loop. So what: memory benchmarks that stop at recall will keep overstating agents, because the expensive failure is not a missing memory but a decision taken without opening it — which is exactly the third check this suite adds.