In the late sixteenth century, news from across Europe flowed into the offices of the Fugger banking house in Augsburg. They wrote about prices, harvests, politics, war. The handwritten sheets that survive in Vienna today are among the earliest examples of a private business intelligence network.1
Speed gave the Fuggers an immediate edge. Their correspondents often knew what had happened before the wider market did. The archive became more valuable with age. Reports could be compared, sources judged and patterns recognised. Each new letter added something to the meaning of the letters already on file.
Companies now hold archives the Fuggers could scarcely have imagined. Email, meeting transcripts, client records, policies, tickets and code preserve much of what an organisation has seen and done. Yet most of that knowledge remains difficult to use. The files survive. The context around them fades. Daily work still depends on people who remember which source can be trusted, which decision superseded another and why an exception was made.
The widening gap between storage and recall sits behind a category that Y Combinator calls the Company Brain. In its Summer 2026 Requests for Startups, partner Tom Blomfield argued that rapid progress in model capability had exposed the next constraint: knowledge specific to each business, scattered across systems and held together by human memory.2
Whatever one thinks of the name, an AI agent cannot work reliably inside a company until it understands how that company works.
From storage to recall
Enterprise search has improved enormously. Given the right permissions, it can retrieve a policy, a meeting note or a client record in seconds. Recall asks for something more: a current account of what the organisation knows, supported by evidence and sensitive to change.
A policy may still be searchable long after it has been replaced. A project decision may appear in several documents, each written at a different stage. Two teams may use the same term for different things. Retrieval finds the material. Useful memory has to establish what is still valid.
Search can find the policy. Memory can tell you whether it is still in force.
Making that distinction requires structure. People, companies, decisions and documents need stable identities, and the relationships between them must survive the meeting or message in which they first appeared. New evidence should update the relevant records while earlier claims remain traceable to their sources. Time matters as much as content: what was true in March may be wrong in September.
Recent research is beginning to treat this as a distinct problem. A University of Illinois paper released on 20 August 2026, “Can Agent Memory Systems Track Evolving State?”, tested whether agents could follow the current state after facts, constraints and decisions changed across several sessions. Existing memory systems, retrieval systems and long-context models were all tested. All of them struggled. Explicitly representing which information had been superseded, and which other facts depended on it, improved the results substantially. The paper is a recent preprint and has not yet completed peer review.3
Another new benchmark, MemOps, treats memory as a lifecycle of operations: remembering, updating, forgetting and reflecting. Its authors found that current systems remain far from uniformly reliable, particularly when they have to reconstruct a sequence of changes from a long and distracting history. A larger context window can preserve more of that history; the current state still has to be reconstructed. MemOps is also a recent preprint.4
Garry Tan, Y Combinator’s president and CEO, has built a useful prototype of this approach. His open-source project, GBrain, contains more than 150,000 pages covering people, companies, meetings and ideas. It combines ordinary retrieval with a knowledge graph, linking records as new information arrives. When a meeting note mentions a company, the system can update the company’s history, the people involved and any open questions connected to them. The knowledge outlives the document.5
GBrain is also designed to revise its own procedures. Tan describes how, after an early analysis produced three false statements about his family, he added a mandatory fact-check involving several models and made it part of the reusable process. Later analyses inherited the correction. The results are self-reported, but the design choice matters. An error changed future behaviour instead of ending as another post-mortem on file.6
An organisation learns when the next attempt carries the lesson.
When memory changes the work
Blomfield developed the same idea in a May 2026 talk, “How to Build a Self-Improving Company with AI”. His first requirement was to make the organisation legible to AI: record the work, preserve the decisions and document the manual steps that employees usually carry in their heads.7
Google Research is exploring what comes next. Its ReasoningBank system extracts useful strategies from successful and failed agent runs, then makes those strategies available during later tasks. The memory contains operating lessons rather than a transcript of everything the agent did. On web-navigation and software-engineering benchmarks, the researchers report higher success rates and fewer steps.8
A living memory can influence the next action and absorb its result. Work produces evidence, which updates the company’s understanding and informs what happens later. Conventional repositories preserve what was known when a document was written. They were never built for this loop.
Once this loop is reliable, the company brain becomes part of the operating model. It can prepare a decision with the relevant history, identify a contradiction before it causes trouble, or show that an apparently new problem has been encountered before. Over time, less has to be rediscovered.
Reliability becomes more demanding once memory influences action. A poor search result inconveniences a user. A poor memory can travel into a decision, a communication or a transaction, making the consequence of an error much larger.
A Swiss bank must remember differently
“Record everything” does not work here. Information is separated for legal, regulatory and commercial reasons. Client confidentiality, business-line controls, jurisdiction, purpose limitation and retention obligations all determine what may be captured and how it may be used.
Swiss data-protection law already establishes the basic discipline: personal data must be collected for a specific purpose that the individual can recognise, and subsequent processing must remain compatible with that purpose. A company brain remains subject to the same obligations.9
These boundaries complicate the idea of a shared corporate memory. One answer may draw on several sources with different entitlements. The user may see some of them. A model can also reveal restricted information indirectly, allowing a protected fact to shape its conclusion without reproducing the source.
Access control therefore has to follow the query from retrieval through reasoning to the final answer. Checking the final screen is too late. The same discipline must apply when an agent writes information back: a useful summary can still become an unauthorised copy if it is stored in the wrong place.
In a regulated institution, permission is part of the answer.
GBrain’s company mode offers each user a login-scoped view and reports that its read routes were fuzz-tested without cross-user leaks. The project offers useful engineering evidence, although its claims have not been independently audited and fall well short of a banking control framework.
A bank would need stronger guarantees. Each significant claim should retain its source and date. Entitlements should be enforced during retrieval, traversal, synthesis and write-back. Conflicting sources need an explicit order of authority. Agent actions must be auditable, and the system must be able to decline an answer when the evidence or permission chain is incomplete.
Financial supervision is starting to move in the same direction. The Financial Stability Board’s June 2026 consultation, “Sound Practices for Responsible Adoption of Artificial Intelligence”, proposes governance across the full AI lifecycle. It asks explicitly whether existing controls are flexible enough for agentic AI.10 That lifecycle view matters. The risk no longer sits only in a model’s answer. It also sits in what an agent retains, how that memory changes and where it is used next.
Memory also needs an expiry mechanism. Policies change, client circumstances move on and access rights end. Some records must be deleted after a defined period; others may remain as evidence while becoming unavailable for routine use.9 Fluent recall of obsolete or impermissible information is a new form of operational risk.
The Swiss version of a company brain will therefore be more selective than the one imagined in Silicon Valley. Its quality will depend as much on what it refuses to retain as on what it can retrieve.
A trustworthy company brain must know what it may remember and when it is required to forget.
What belongs to the company
Most enterprise AI discussions begin with model selection. That decision matters, although probably for less time than many current programmes assume. Models change. Prices fall. Vendors trade places. A company that binds its operating knowledge to one model risks having to rebuild its memory each time the market moves.
Tan summarises his architecture as “fat skills, fat code, thin harness”: the operating knowledge sits in procedures that people can inspect and edit, while a relatively simple layer routes each task to an appropriate model. Behind the terminology is a simple design choice: keep the company’s operating knowledge portable across models.
The model is rented. The memory is yours.
The durable asset is the accumulated account of how the company works: its sources, decisions, relationships, exceptions, checks and permissions. Competitors can license the same models. They cannot immediately reproduce the context built through years of work.
Preserving that advantage means keeping the memory intelligible outside any particular AI system. Procedures stay readable. Claims lead back to evidence. Permissions sit in enforceable controls rather than prompts. The company must be able to replace the model without losing its understanding of itself.
The Fugger archive became valuable because each report made the next one easier to interpret. Modern companies can create the same compounding effect at a much greater scale. A decision can retain its reasoning. A mistake can alter the next attempt, long after the people involved have moved on. Knowledge survives its original holders without escaping the boundaries that make it trustworthy.
Intelligence will become widely available. Recall will not.
Every company will be able to rent a powerful model. The advantage will belong to those that do not have to learn the same thing twice.
We cannot stop talking about this one. We are putting an evening together.