FluctlightDB: A Memory Model of Data for AI Agents

We argue that long-term agent memory is a third data model—distinct from relational databases and vector stores—with native write semantics (experience: encode, separate, consolidate, provenance) and read semantics (activate: cue-driven hybrid retrieval over a memory graph). We present FluctlightDB, an embedded open-source engine (MIT) exposing experience() and activate() as native API primitives, one durable store per agent. Headline results (reproducible; harnesses and frozen JSON in the GitHub repository): • LoCoMo (10 conversations, 1,982 gold evidence spans): 98.1% mean evidence recall @ k=150 (warm = cold-start identical) • BEIR SciFact: nDCG@10 0.645, matching Chroma + MiniLM at equal latency • FAMB agent-memory suite: 97–98% macro Software, benchmark harnesses, and LaTeX source: https://github.com/voxmastery/FluctlightDBPreprint PDF attached. arXiv cs.DB submission pending.