Agent memory research code: a third path between "absorb everything into the weights" and "store everything in an external retrieval index."
Most agent-memory designs assume one of two extremes: consolidate experience into model weights (slow, destructive, opaque) or keep an ever-growing external store and retrieve from it (cheap to write, hard to keep predictive). This project studies an old and under-used alternative from the online-learning literature: maintain several local predictive models simultaneously, update all of them online as the stream arrives, and at each decision point act on whichever currently predicts best. In a world too big to model globally, selection over cheap local models can beat both extremes — and the experiments here measure when and why, including a full evaluation on MemoryAgentBench.
artifact/— the frozen reproducibility package, preserved byte for byte:code/— every generator and analyser behind the reported numbers (stream construction, matched-budget baselines, non-stationarity and abstention experiments, follow-the-leading-history traces, agent evaluation, figure generators)CLAIM-*.md— analysis plans written before the runs and shipped unchangedCORRECTIONS.md— every correction and retraction made during the workMANIFEST.txt+verify.sh— SHA-256 of every shipped file and a checker
cd artifact
sh verify.sh # confirms every file matches MANIFEST.txt
cat README.md # per-experiment entry points and environment notesThe artifact/ tree is exactly the package the reported numbers were generated from;
it is intentionally not edited here (verify with sh verify.sh). Anything at the
repository root is packaging only.
Research code release. A paper reference will be added here once available.
MIT (see LICENSE).