Prelaunch preview. Explore AROS products and skills. Checkout is in sandbox; no real payment.
Blog

Beyond the chat window

Scientists work with AI by conversation, which loses the thread; private systems keep it but wall it in. We are building a shared research-experience system.

A chat log is not executable, inspectable or shareable; research with AI needs an environment that is all three.

Wang Ziyi14 min readEnglish

Most researchers who work with AI today, in biomedicine especially, do it through nothing more than a chat window. A question, an answer, a correction, another question, for hours. The more experienced build something around the window: for example prepared prompt files, a tree of notes the model is told to read first, a home-made memory store, a workflow runner. Yet faced with a tool that forgets everything between sessions, both habits are reasonable responses. Both also fail in some ways that matter to the research community.

The conversation is inefficient because it cannot hold a method, and because it cannot complete a long chain of tasks by looping on itself. The private system is inefficient because it serves one person. In the previous essay we argued that the rules for auditing research (peer review, literature synthesis, meta-analysis, the institutions that manage and audit research funding) should run as software while the work is being done. The software nexitia.ai is building to run those rules has two further properties a chat log lacks: a person can read it, and a person can hand it on, to other people and to AI agents. This essay is about the environment that gives a scientist and an agent those properties together: memory the scientist can consult, trace and share, workflows the scientist can see, and a way to share methods and paradigms without sharing data. We are building it as one desktop app, Nexitia.

What a conversation cannot hold

A chat is an append-only record in which everything has the same weight. The method that worked on turn 47 sits next to the false start on turn 12, the pasted table, the apology and the retry. Nothing in it is versioned, nothing is tested, and nothing marks which lines the eventual result depends on. When the session ends, the researcher is left with a transcript to mine or a summary the model wrote about itself.

The models do not hold the thread well either, which is called attention loss. When Laban and colleagues took tasks that models solve from a single, fully specified instruction and revealed the same instruction one piece per turn through a simulated user, the models made assumptions early, built on them, and rarely recovered once a conversation had gone wrong [1]. Across six tasks, the 15 models they tested scored 39% lower on average than when given the full instruction at once [1]. Even within one long prompt, where in the context a fact sits changes whether a model uses it: Liu and colleagues found that accuracy is highest when the relevant document is at the beginning or the end of the input and lowest when it is in the middle, where one model did worse than with no documents at all [2]. A long scientific conversation is the worst case for both findings: the decisive detail is usually in the middle, stated once, many turns ago.

Private systems answer the first problem and create a second one. A researcher who keeps a prompt library and a memory store has something a conversation cannot give them: continuity. But the store is written in their own vocabulary, structured around their own projects, and readable by no one else. Science has met this wall before, without AI. When Collins interviewed twelve laboratories building TEA lasers in 1971 and 1972, none had managed to build one and make it work from written sources alone; every laboratory that got a working laser had learned by visits, telephone calls or staff moving from a laboratory that already had one [3]. The method travelled through real, physical contact between people, not through the record. Four decades later, with a journal policy requiring data and code, Stodden and colleagues could obtain the artefacts for only 89 of 204 computational Science papers, and estimated that the results could be reproduced for only about 26% of the 204 [4]. A private AI memory is the same wall with a newer surface: the know-how exists, it is even written down, and it still does not travel.

Memory the scientist can open

Nexitia brings three pieces into one window: Nexitia Core (listed today as AROS Core), the memory and orchestration engine; the IDE, where a person inspects context, files and outputs; and Commander, which coordinates work across a laboratory's machines. The chat window stays, beside the canvas rather than in place of it. Nexitia Core is the memory our tools and agents work from. It keeps more than one kind of knowledge apart, because they are used differently: observations tied to an event, reusable facts, conceptual or heuristic models, and procedures that guide action (white paper, section 9). A conversation flattens these into one stream; Nexitia Core keeps them as what they are, with their origin, version and scope, so that a later task can ask for a validation procedure without inheriting the assumptions of the project that wrote it.

Figures from our own reference instance show what that memory has to do. As of 7 September 2026 it held 22,714 facts from experiments on the physical world, 16,591 episodic experiences and 2,197 specialised skills in its skill hierarchy, the brain of one laboratory (shared-brain statistics). Prompts had pulled in only 282 distinct skills, 13% of the library, across 618 recorded retrievals. We show that unused 87% on the public page because it is the argument for curation over accumulation. A bigger memory is not a better one; a memory is useful to the extent that the right piece reaches the agent at the right moment, and that the person can see which piece it was.

That last requirement, that a person can see which piece the agent used, is a system feature we are developing. Nexitia Core loads context by progressive disclosure: an agent starts from compact descriptions of candidate tools and workflows, reviews inputs, assumptions and what has been verified, loads the selected operation's instructions and checks, and opens detailed methods and references only when needed. The researcher inspects the selected context and its sources; software, not the model, checks permissions, versions and input compatibility. The memory an agent reasons from is a view the researcher can open and review: the facts, knowledge items, skills and wiki pages it used, and the links between them, drawn as a graph rather than buried in a transcript. In Nexitia that view uses the same canvas as CohortDAG and Tracyte: steps grouped into phases, phases collapsed to a single node at a distance and opened when the researcher looks closer, so that a project of hundreds of steps reads at the altitude the person needs.

The agents' own work is drawn the same way. A conversation in the Nexitia chat runs as a job, so what it was given and what it produced enter memory like any other run instead of disappearing with the window. When several agents work on a task, the task appears as a workflow: each agent's steps are nodes, and when an agent uses one of our AI-native tools, say Tracyte for a single-cell RNA-seq analysis, that use appears as a node of its own, which opens the tool at the session and step the agent touched. The researcher reviews the agents the way they review a pipeline, by looking at it. Agents propose; a person decides, in the tool or in Nexitia, before anything is written.

The memory also edits itself, and the edits are on the record. Our reference instance keeps a lifecycle ledger of every registration, rewrite and consolidation: 1,037 events by 7 September 2026, including 828 that merged a component with a duplicate and 175 were rewrites by the prompt optimiser (shared-brain statistics). A system that improves its own skills needs that ledger for the same reason a laboratory needs a notebook: so that a person can see what changed, when, and by whose hand.

A tool is the unit a person can inspect, which is why related work is split into separate AI-native tools, each with its own versioned skills and scientific context, and Nexitia Core organises them by field and workflow similarity with cross-links for shared methods (white paper, section 11). Each tool stays a product of its own and plugs into Nexitia as a node. A single assistant that does everything shows the researcher nothing; a tool with a canvas shows them the graph.

Share the method, keep the data

The wall around a private system comes down only if a method, or an experience, can leave one laboratory without the laboratory's data leaving with it. A federation of experience is how we do that, and it is built as two separate paths rather than one switch (privacy policy, section 4).

The first path is an organisation's own brain, synchronised for the laboratory and private to it. The second is the commons, and it carries procedural knowledge (skills, knowledge items, workflows and policies) along with mental models and insights that pass a screen. Episodic experiences, facts drawn from experiments or summaries about the physical world, and the links of the memory graph never leave the instance. Before anything is contributed, rows containing file paths, web addresses or e-mail addresses are removed (de-identification), home directories in skill text are rewritten, and attached artefacts pass an allow-list that keeps out sensitive material such as privacy-related row-level tables, sequencing files or medical images. The screen catches syntactic leaks; what a paragraph means is the contributor's judgement, so the contributor chooses what goes into the synced and shared commons or into the laboratory's own shared environment.

What the commons publishes about itself follows the same rule. Federation-wide figures include only users who consented to publicity, and are released only once at least five laboratories are actually contributing, applying the k-anonymity principle that no release should let a contributor be told apart from fewer than four others [5]. Below that floor nothing is published, because a total across two or three laboratories can be unpicked by one of them. No per-laboratory or per-machine breakdown is ever published either, since an instance label is a hostname and would re-identify its contributor (shared-brain statistics).

Reading does not require contributing. The curated public procedural bank is open through the Model Context Protocol at no charge: eight read-only tools, an organisation key, and admission caps so that one heavy client cannot crowd out the rest. A researcher who has never synchronised anything can point their own agent at the methods others have chosen to share.

Collins's laser builders needed a visit because the record held the result and not the reasoning. A shared skill in the commons carries the method, its context and the human reasoning and experience behind its important choices, and keeps its origin and version as it is tested and adapted elsewhere (white paper, sections 13 and 14). That is the difference between a federation of experience and a bigger catalogue of local projects.

An exchange for workflows

A workflow a researcher builds with their agent in CohortDAG, Tracyte or any future nexitia.ai product, with or without AI nodes and with the share of AI set by the researcher (the AI dial of the first essay), is a method that someone in a neighbouring field could run tomorrow. The last part of the environment is the curated exchange that lets this happen smoothly. A contributor may keep a workflow private, share it on terms they choose, or offer it as a maintained workflow package; every package names its author, intended use, dependencies, evidence, licence and update policy, and publication requires review and documented supporting data (white paper, section 19). Compensation is designed to reward maintained usefulness rather than a download, because a method others rely on creates work: compatibility, questions, revisions.

The same verification records that make a node trustworthy inside one laboratory are what make it worth exchanging between two. A workflow arrives with the reference it was checked against and the items it matched; a laboratory elsewhere runs it against its own requirements and, if it finds the exception that makes the method clearer, contributes the adaptation back.

What comes next

Three claims above are testable, and we are testing them. We are measuring how much of a paused project an agent can resume from Nexitia Core's memory alone against the same agent given the chat transcript. We are measuring how often a workflow shared through the commons runs unchanged in a second laboratory, and what the second laboratory has to change when it does not. And we are testing whether the sync boundary withholds everything a contributor marks as private while the commons screen catches a set of deliberately planted syntactic and file leaks. The results will appear here.

References

  1. Laban P, Hayashi H, Zhou Y, Neville J (2025). LLMs get lost in multi-turn conversation. arXiv:2505.06120 (preprint). In more than 200,000 simulated conversations in which a GPT-4o-mini "user" revealed one piece of the instruction per turn, 15 models scored on average 39% lower than with the full instruction in one turn, across six tasks; aptitude fell 16% and unreliability rose 112%.
  2. Liu NF, Lin K, Hewitt J, Paranjape A, Bevilacqua M, Petroni F, Liang P (2024). Lost in the middle: how language models use long contexts. Transactions of the Association for Computational Linguistics 12: 157–173. doi:10.1162/tacl_a_00638. In multi-document question answering over 20 constructed retrieval passages, mid-2023 GPT-3.5-Turbo answered 75.8% of questions when the relevant passage came first, 53.8% when it was tenth and 63.2% when it was last, against a closed-book score of 56.1% with no passages.
  3. Collins HM (1974). The TEA set: tacit knowledge and scientific networks. Science Studies 4(2): 165–185. doi:10.1177/030631277400400203. Interviews at seven British and five North American TEA-laser laboratories in 1971–72: no laboratory Collins spoke to had built a working laser from written sources alone, and every successful one had learned by visits, telephone calls or transfer of staff from a laboratory that already had one; Collins notes the field is untypically suitable for the thesis.
  4. Stodden V, Seiler J, Ma Z (2018). An empirical analysis of journal policy effectiveness for computational reproducibility. Proceedings of the National Academy of Sciences 115(11): 2584–2589. doi:10.1073/pnas.1708290115. Of 204 computational Science papers published after the February 2011 data-and-code policy, artefacts were obtained for 89 (44%: 24 already linked and 65 shared by the 180 authors e-mailed), and 21 of 22 papers drawn at random from the 56 judged reproducible on inspection were reproduced, giving the authors' estimate of 26% for the full sample.
  5. Sweeney L (2002). k-anonymity: a model for protecting privacy. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 10(5): 557–570. doi:10.1142/S0218488502001648. Defines k-anonymity: each released record is indistinguishable on its quasi-identifiers from at least k−1 others, the principle our five-laboratory floor applies to published totals.

Our records:

  • Shared-brain statistics, the federation page, 7 September 2026 snapshot of one laboratory's reference instance: 22,714 world facts, 16,591 episodic experiences, 2,197 skill-hierarchy rows; 282 distinct skills retrieved across 618 retrievals (13%); lifecycle ledger of 1,037 events, 828 merged, 175 rewritten by the prompt optimiser; publication rules (publicity consent, five-lab floor, no per-lab or hostname breakdown).
  • Nexitia Core (listed on the products page as AROS Core): the memory and retrieval foundation; discover, select, execute, investigate; the researcher inspects the selected context and its sources; permissions, versions and input compatibility checked in software.
  • The AROS white paper, edition 30 October 2026, revision 2: section 9 (kinds of knowledge), section 11 (tools with their own skills and context, organised by field and workflow similarity), sections 13 and 14 (a shared artefact carries the method, its context and its reasoning; the Global Scientific Brain retains identities, contexts and versions), section 19 (what each offering names).
  • Privacy policy, 17 October 2026, section 4: organisation sync and commons contributions are separate paths; the exporter's filters and screening do not guarantee that all sensitive content is excluded; what enters a snapshot is determined by the local instance.
  • MCP server: Standard Read Access at no charge, eight allowlisted read-only tools, an organisation API key with the read scope, admission caps.
  • Pilot and pricing: the curated exchange of authorised tools, skills and workflows; contributors may keep work private, share under chosen terms or propose a licensed package; publication requires review and documented rights.