Alibaba Cloud's project hub published the write-up on 28 October. Quen, a Vietnamese word meaning to forget, chosen because it sounds like Qwen, is a memory layer built by solo developer Pham Thanh Hang on the premise that append-only memory eventually asserts stale facts as current. On the write path, Qwen extracts subject-relation-object triples, a salience gate skips facts a base model already knows, and confidence is derived from source authority. A background dream pass consolidates episodic memories, resolves contradictions and evicts what has decayed; deletions leave tombstones so forgetting stays auditable. Each memory carries FSRS spaced-repetition retention state, bi-temporal validity and a confidence score. At answer time, low-trust memories are checked against the live source: confirmed raises confidence and logs a good review, refuted tombstones the memory and logs a failure, unverifiable forces the answer to hedge. Contradictions in functional relations are resolved first by a deterministic slot rule with no model call, with Qwen natural-language inference as a fallback. Reported results: 0.933 against 0.400 for append-only RAG on the team's staleness probe, and 0.423 against 0.253 on LongMemEval-S. All model and embedding calls run on Qwen through Alibaba Cloud Model Studio.
Our takeFAMA is defined by the team, which makes 0.933 a claim about their probe as much as about the system; the comparison still holds because both systems face the same probe. The engineering worth borrowing is the deterministic contradiction rule: resolving same-subject, same-relation conflicts without a model call keeps memory auditable. The cost is latency and a hard dependency on the source being reachable at answer time. A system that must re-read the web to speak is a system that goes quiet when the web does.