AWhat changed
Two memory systems can read the same text yet preserve different details. MIRAS makes four design choices explicit: the memory architecture, the objective used to learn associations, the rule for retaining old information and the learning algorithm. The researchers use this perspective to construct Moneta, Yaad and Memora, including variants based on robust losses and different regularizers. Their experiments compare language modeling, recall and other tasks. The contribution is this organizing framework and its demonstrated new variants, rather than invention of associative memory or the optimization tools it combines.
Scientific scope
Two memory systems can read the same text yet preserve different details. MIRAS makes four design choices explicit: the memory architecture, the objective used to learn associations, the rule for retaining old information and the learning algorithm. The researchers use this perspective to construct Moneta, Yaad and Memora, including variants based on robust losses and different regularizers. Their experiments compare language modeling, recall and other tasks. The contribution is this organizing framework and its demonstrated new variants, rather than invention of associative memory or the optimization tools it combines. The paper reports bounded experiments, with different variants winning different metrics. Its chunkwise parallel updates are not identical to arbitrary tokenwise nonlinear learning. Robust losses do not guarantee immunity to every outlier, and numerical results do not certify every mathematical interpretation in the paper. Existing Titans, fast-weight, Huber-loss and mirror-descent contributions remain separate. The neural model named Memora here is distinct from Xia and colleagues’ later external MEMORA system.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The paper reports bounded experiments, with different variants winning different metrics. Its chunkwise parallel updates are not identical to arbitrary tokenwise nonlinear learning. Robust losses do not guarantee immunity to every outlier, and numerical results do not certify every mathematical interpretation in the paper. Existing Titans, fast-weight, Huber-loss and mirror-descent contributions remain separate. The neural model named Memora here is distinct from Xia and colleagues’ later external MEMORA system.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.9. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction