Pathan (2602.22152) introduces Stream Neural Networks — an architecture designed for irreversible information streams where you can't replay the data.
Note (2026-11-11): checked against the paper's full text (this essay had been logged as held up on an abstract-level read). The paper is by Amama Pathan (independent research) and is correctly described as introducing stream neural networks for irreversible input streams, each input processed exactly once, with persistent state and three structural guarantees (collapse of stateless mappings, bounded state, contraction for a decay parameter below 1). The essay's account of learning goes beyond it. The paper says the stream neuron architecture 'defines execution behavior only' and that learning rules, memory management and stability enhancements operate independently of the neuron's temporal update and are addressed in subsequent work; the behaviour it reports 'does not depend on learning rules, optimization procedures, or parameter adaptation'. So the essay's statements that learning happens in a single forward pass, that no backpropagation through time or replay buffer is needed, and that compression of the past is part of the learning describe things the paper leaves to later work (I found no mention of compression). The essay's separation of fast adaptation from slow integration, mirroring working memory and long-term potentiation, is not in the text, which says the system does not model biological spiking behaviour. The paper's evidence is a phase-space analysis and continuous tracking experiments; I found no benchmark comparison with multi-epoch training, which the essay presents as the practical test.
Most neural networks assume you can iterate over the training set multiple times (epochs). This implicitly requires reversible computation: access to the past. But real-world environments — sensor feeds, market data, conversation — present information as a one-way stream. Once a datum passes, it's gone. Conventional architectures trained on streams degrade into reactive filters: they respond to the current input but lose coherence over long horizons because they have no mechanism for accumulating temporal structure without rehearsal.
Stream Neural Networks maintain persistent temporal state that evolves with each input but never requires replaying old inputs. The key departure: learning happens in a single forward pass through time. No backpropagation through time, no experience replay buffer, no stored trajectories. The network's state carries a compressed summary of everything it has seen, and the compression is part of the learning — not a separate memory system bolted on.
The architecture separates fast adaptation (responding to the current input) from slow integration (updating the persistent state that shapes future responses). This mirrors the biological distinction between working memory and long-term potentiation, though the mechanism is different.
The practical test is whether single-pass learning can match multi-epoch training on benchmarks where multi-epoch is possible. If it can, the stream paradigm is strictly more general — it works everywhere epochs work, plus environments where epochs are impossible. If it can't, the question becomes how large the gap is and whether the gap matters less than the ability to operate in truly streaming environments.