A new paper proposes a dedicated memory layer for AI agents, and reports its results under three different names for what should be one system.
FluctlightDB is an embedded engine that treats long-term agent memory as its own data model, separate from relational or vector search, with write functions for encoding and consolidation and a read function called activate() for cue-driven recall. On LoCoMo, a 10-conversation benchmark with 1,982 gold answer spans, the authors report 99.0% evidence recall, credited to something called CHORUS. On LongMemEval-S, a 500-question benchmark, a retrieval harness scores 97.6% session recall and an end-to-end QA stack scores 97.4%, with no system name attached at all. On BEIR SciFact, a system labeled CHORUS/PRISM narrowly edges the Chroma vector database on two ranking metrics. A small in-house regression suite the authors built themselves scores 100%.
The underlying pitch is sober by AI-paper standards: agent memory needs cue-driven, provenance-aware recall that neither relational databases nor vector search was designed for, and the authors say plainly they are not trying to beat Mem0, Zep, or HippoRAG, just sit underneath them. But a paper introducing one new system should use one name for it. Whether FluctlightDB, CHORUS, and CHORUS/PRISM are the same engine, different versions, or different components is never explained, and that gap sits right next to headline numbers the paper wants readers to trust.
The code and frozen benchmark JSON are MIT-licensed and a pip install away, so anyone can rerun the numbers. They just cannot yet tell you, from the paper alone, what to call the thing they are testing.