Round three
This round reviewed twelve publications on memory, replay, compositional reasoning, information sampling, and continual learning. The question was simple: what has actually been established about how the brain organises experience, retrieves what matters, and recombines it into new answers, and what remains unproven.
Human replay reorders experience by learned relationships rather than by the order it was encountered (Liu and colleagues, 2019). Which memories get replayed at all appears to be prioritised by how useful they are likely to be for the decision at hand (Mattar and Daw, 2018), and replay can occur without predicting the next action, which suggests memory functions separate from immediate planning (Gillespie and colleagues, 2021).
Hippocampal ripple activity is associated with compositional inference, recombining familiar elements into candidate solutions (He and colleagues, 2026). Entorhinal and ventromedial prefrontal cortices generalise the structure of a task and apply it to new situations, retrieving earlier cases by relationship (Baram and colleagues, 2021).
Humans generalise systematically to new combinations of familiar rules. Standard neural networks do not, unless they are explicitly meta-trained on recombination (Lake and Baroni, 2023). Structure reuse is a capability with a cost, not a default.
In simple choice, where people look reflects near-optimal information sampling: the next piece of information examined is the one expected to resolve the decisive uncertainty first (Callaway and colleagues, 2021). Analysis order is chosen, not fixed.
Networks that keep training on new data progressively lose the capacity to learn new rules, a distinct failure from forgetting old ones (Dohare and colleagues, 2024). Relation-aware memory structures improve on standard retrieval for accumulating knowledge over time, though extraction errors still dominate outcomes (Gutiérrez and colleagues, 2025; Wu and colleagues, 2025).
Evidence on scaling agent systems indicates that adding coordination helps when the task's structure calls for it and can hurt otherwise, with reported improvements varying substantially by task (Kim and colleagues, 2026 preprint).
These are research foundations and comparison methods. Their authors have not evaluated or endorsed Encevra. Each entry carries the limitation we are keeping in view.
Recombining familiar elements into candidate solutions.
Limitation: Intracranial observations in 28 patients. We reviewed the original abstract only.
Read the sourceReorganising experience by learned relationships rather than by order seen.
Limitation: MEG evidence in humans. It does not establish the AI mechanism we propose.
Read the sourcePrioritising what to reconsider by how useful it is likely to be.
Limitation: Normative simulations, not a complete neural implementation.
Read the sourceTesting memory functions separately from immediate planning.
Limitation: A rodent task in which replay did not predict the next choice.
Read the sourceRetrieving earlier cases by relationship, then checking the differences.
Limitation: A trained laboratory task with fMRI, not open-ended business work.
Read the sourceTesting on new combinations of familiar rules, not memorised cases.
Limitation: Human instruction tasks and specially trained models.
Read the sourceAsking the question that resolves the decisive uncertainty first.
Limitation: An approximate model of simple choices, with known mismatches.
Read the sourceComparing relation-aware retrieval against standard retrieval.
Limitation: Benchmark-specific evidence. Extraction errors still dominate outcomes.
Read the sourceMeasuring capacity to learn new rules separately from retaining old ones.
Limitation: The proposed method does not itself solve forgetting.
Read the sourceScoring claim support and citation coverage as separate measures.
Limitation: Citation metrics do not certify that any individual decision is correct.
Read the sourceTesting time reasoning, updates, multi-session recall, and abstention.
Limitation: Conversational histories differ from business document sets.
Read the sourceMatching the amount of coordination to the structure of the task.
Limitation: A preprint. Reported improvements vary substantially by task.
Read the sourceEncevra's research draws on cognitive science and computational neuroscience to investigate how AI systems can organise evidence, retrieve relevant experience, compare alternatives, and revise conclusions. Published findings guide our experiments. Encevra's own capabilities have to be established by evaluating the product itself.
