Нейросеть сама открыла метод «дворца памяти»: исследование Nature Machine Intelligence

Researchers led by Li (Li et al.) published a paper in Nature Machine Intelligence on July 20, 2026, showing that recurrent neural networks trained on the free recall task independently discover several human-like memory strategies at once — and the most accurate of them use an index-based mechanism reminiscent of the ancient "memory palace" mnemonic technique.

What the study found

Recurrent neural networks (RNNs) optimized for the free recall task were found to use not one but several different memorization strategies. According to Nature Machine Intelligence, these strategies go beyond the classical temporal context models that have for decades been considered the primary explanation for how humans recall lists of words in arbitrary order.

Free recall is a classic memory test: a person is shown a list and then asked to reproduce it in any order. For a long time, human behavior on this test was described specifically by temporal context models, in which the probability of recalling a word depends on how close in time it was presented to its neighbors. The new study shows that this task does not have a single optimal solution.

*Publication — July 20, 2026, Nature Machine Intelligence, DOI 10.1038/s42256-026-01274-0

*Authors — Li et al.

*Method — recurrent neural networks (RNNs) optimized for free recall

*Key finding — top-performing models use an index-based mechanism, an analogue of the "memory palace"

*What they outperformed — classical temporal context models

What the "memory palace" has to do with it

The most effective models relied on an index-based mechanism: they assigned the elements to be remembered a kind of address and accessed them via these markers, instead of scanning through the entire sequence. The trick resembles the "memory palace" technique (the method of loci) used by ancient orators — facts are mentally "placed" in the rooms of an imagined building and then retrieved by walking a predetermined route.

Importantly, the network was not given this strategy ready-made. It arrived at it on its own, simply by optimizing for recall accuracy — meaning the index-based scheme turned out to be an advantageous solution to the task itself, not a trick built in beforehand.

Why it matters

The work brings artificial and biological memory mechanisms closer together. If a neural network arrives at a strategy resembling human mnemonics without any hint, this is an argument that such strategies are an optimal solution to the recall task, not merely a cultural invention. For neuroscience, this is a testable hypothesis about how the brain organizes free recall.

For AI, the question of long-term memory remains open: language models keep context within a limited window and organize old memories poorly. The strategy in which elements are addressed by index rather than reviewed sequentially echoes the way modern systems bolt on external databases and vector search. The study by Li et al. gives this a biologically plausible grounding.

Top-performing models use an index-based mechanism reminiscent of the "memory palace" technique. — from the paper by

Li et al., Nature Machine Intelligence (DOI 10.1038/s42256-026-01274-0)

What this means

Neural networks increasingly serve not only as applied tools but also as models of the brain. By optimizing a network for a specific memory task, researchers obtain testable hypotheses about how human recall works — and hints for future AI systems with more structured long-term memory.