FluctlightDB: A Memory Model of Data for AI Agents
Preprint · June 2026 · DOI: 10.5281/zenodo.20949890 · arXiv pending
Author: Ganesh S · ORCID 0009-0006-7758-4114 · voxmastery@gmail.com
One-line claim
Long-term agent memory is a third data model — not SQL rows, not vector ANN alone. FluctlightDB is an embedded engine with native experience() / activate() semantics.
Headline results
| Benchmark | Metric | Result |
|---|---|---|
| LoCoMo (10 conv, 1,982 gold spans) | Mean evidence recall @ k=150 | 98.1% |
| BEIR SciFact | nDCG@10 | 0.645 (ties Chroma + MiniLM) |
| FAMB | Macro (index / agent) | 98% / 97% |
Frozen metrics: fluctlightdb-benchmarks
Abstract
For fifty years, data systems answered two questions: which records match a predicate (relational), and which vectors lie nearest a query (vector). Autonomous agents ask a third: what have I learned, and what of it can I trust?
We present FluctlightDB, an embedded brain-native database with write path experience() and read path activate(). On full LoCoMo it recalls 98.1% of gold evidence (warm and cold-start identical). On BEIR SciFact it matches a tuned Chroma baseline; on FAMB it scores 97–98% macro.
Install
pip install "fluctlightdb[native]"
from fluctlightdb import connect
brain = connect("/tmp/agent-brain")
brain.experience("User prefers dark mode", context="settings", salience=0.8)
print(brain.activate("theme preference"))
brain.checkpoint()
Links
| Resource | URL |
|---|---|
| DOI (Zenodo preprint) | https://doi.org/10.5281/zenodo.20949890 |
| LaTeX source | https://github.com/voxmastery/FluctlightDB/tree/main/papers/arxiv-v1 |
| Interactive viewer (Space) | https://huggingface.co/spaces/Voxiesz/fluctlightdb-paper-viewer |
| GitHub | https://github.com/voxmastery/FluctlightDB |
| PyPI | https://pypi.org/project/fluctlightdb/ |
| Venue plan | https://github.com/voxmastery/FluctlightDB/blob/main/docs/RESEARCH_VENUES.md |
| Reproduce benchmarks | https://github.com/voxmastery/FluctlightDB/tree/main/benchmarks |
Citation
@article{s2026fluctlightdb,
title={FluctlightDB: A Memory Model of Data for AI Agents},
author={S, Ganesh},
year={2026},
doi={10.5281/zenodo.20949890},
url={https://doi.org/10.5281/zenodo.20949890},
note={Preprint. Software: https://github.com/voxmastery/FluctlightDB}
}
See also CITATION.cff on GitHub.
Metric note
LoCoMo evidence recall = fraction of gold dialogue evidence in retrieved context (official RAG metric). Mem0/Zep often report LLM-as-judge end-to-end QA — a different, harder number. Do not compare 98.1% recall to ~92% QA without naming the metric.
License
MIT — engine, harnesses, and this paper draft.