A Proposal to Connect AI Laboratories Through Shared Scientific Memory

The scientific memory wall

In Evolving Scientific Intelligence: From Fragmented Evidence to Distributed Experimental Loops, Yuheng Ding of the Hong Kong University of Science and Technology argues that scientific AI faces an evidence bottleneck. Computational capabilities are growing, but experimental information remains fragmented, delayed or unpublished. Ding calls this mismatch the “scientific memory wall.”

Research papers communicate findings and assign credit, but they rarely preserve every exploratory trial, negative result or unexplained observation. According to the preprint, this filtering can leave models with a distorted picture of when experiments succeed and fail. Even results eventually published may arrive too late to inform another laboratory’s next decision.

Ding’s proposed response, Evolving Scientific Intelligence, or ESI, would complement publication with shared, machine-readable experimental memory. It is an architectural proposal drawing on existing research, rather than a report of a completed network delivering new discoveries.

Keeping evidence useful beyond its original purpose

ESI would preserve measurements alongside experimental conditions, procedures and provenance: the information needed to trace how a result was produced. Failed trials and anomalies would remain available even when their future relevance is unclear.

The proposal separates a record’s immediate usefulness from its long-term value. A reaction study focused on final yield, for example, might later support an investigation of reaction speed if its original time-course measurements survive.

Storage alone is insufficient. Ding advocates shared scientific meanings and structured fields, while retaining representations suited to different data types, such as molecular graphs or protein sequences. Records would connect four elements: the decision to run an experiment, its evaluation method, the resulting evidence and any representation supplied to a model.

Complete raw records could remain at local laboratories, with coordinators receiving relevant results and retrieving details when needed.

Turning connections into experiments

The preprint calls the recombination of existing evidence into new hypotheses “information collision.” Shared variables, mechanisms or structural analogies could help models identify connections across disciplines. Those connections would generate proposals, not establish that the proposals are correct.

A hypothetical example connects battery research with enzyme electrodes. A battery laboratory rejects a quinone molecule because its solubility is too low for energy storage. Another laboratory needs a molecule that transfers electrons between an enzyme and an electrode, potentially at a different concentration.

ESI could recognize that rejection for one application does not rule out usefulness in the other. It would then organize tests of the molecule–enzyme pairing, including controls and independent measurements. This scenario illustrates the intended workflow; it is not an experimental result reported for ESI.

Choosing both the candidate and the test

Ding argues that deciding what to investigate should also involve deciding how to evaluate it. Computational screening may be cheaper than laboratory work, but unreliable screening can waste resources on false positives or exclude valuable candidates through false negatives.

ESI therefore would not impose a universal sequence in which every candidate must pass a simulation before physical testing. It would select candidates and evaluation methods together, considering scientific value, uncertainty, reliability and cost.

Physical measurements would also help recalibrate predictive evaluators. Revised evaluators should be checked against experimental results not used for their update. The paper cautions against allowing agreement within an AI generation-and-evaluation loop to substitute for independent evidence.

Three levels of feedback

The architecture separates responsibilities across three levels. L1 handles local sensing and experimental execution. L2 generates hypotheses, chooses candidates and evaluators, and coordinates work across laboratories with different capabilities.

L3 provides slower oversight of objectives and decision criteria. Its role is to notice when a system becomes better at optimizing a score without improving the outcome that matters. In the hypothetical electrode example, increasing initial current could undermine accurate glucose measurement over repeated use. L3 would redirect optimization toward sustained measurement accuracy.

What remains unresolved

The proposed benefits depend on problems the preprint explicitly leaves open. Common formats do not automatically give models meaningful cross-domain understanding. Combining numerical measurements also requires compatible definitions, units, conditions and calibration; otherwise, equipment differences could appear to be scientific patterns.

Complete records do not guarantee trustworthy results, and independent verification may consume the resources that sharing was meant to save. Credit, payment and incentives for contributing negative results remain unresolved too.

Ding calls the collective infrastructure effort an “Information Manhattan Project.” Its central ambition is evidence-driven discovery across laboratories, but the paper leaves both the mathematics of discovery-chain propagation and practical arrangements for sustaining participation as open challenges.