The company brain is becoming the company. The unsettled question is who gets to keep it.

Memory is where the value, and the dependency, now accumulate

For a year the field has raced on one axis: how accurately can a system recall the right fact at the right time. That race leaves out the part that decides whether a company brain can be trusted with anything that matters: who is allowed to see a fact, where it came from, and whether it is still true.

A company brain compounds. Every decision and correction that lands in it makes the next answer better, and the more of the firm's judgment lives there, the more that layer decides what the business can prove and how easily it could walk out the door inside someone else's platform. Most of what ships today is strong on recall and thin on exactly that: whether a fact is current, whether the asker was allowed to see it, and where it came from. Tolerable in a demo. A liability the moment the memory drives real work under a regulator's gaze.

The market just agreed the loop is the asset. It has not agreed you can keep it.

In the space of a month, the largest names in the industry converged on one point. The model is becoming a commodity. The durable asset is the learning loop a company builds on top of it: its institutional memory, its private evals, the accumulated judgment that gets a little better every time it is used. Agreeing that the loop matters is a different thing from letting a company keep it.

• Microsoft Frontier Company, 2 July. A $2.5B operating unit, 6,000 engineers embedded at customers, pledging that "a customer's IQ is protected." The learning loop turned into a business.
• Palantir, 30 June and 7 July. "Data retention is your treasure. Transfer it at your own peril," then a fifteen-step sovereignty playbook two weeks later that arrives at our architecture one layer up: model-agnosticism, a signed audit trail, an ontology that must "sit outside the model itself." When the maximalist incumbent converges on this shape, it is market consensus, not a Lexenne opinion.
• Google OKF, 12 June. An open, Apache-2.0 format for organizational knowledge at rest. The shape of a company's knowledge becomes a commodity, which helps everyone building on it.

None of them answers the question that decides who benefits: can a company walk out with the loop it built and stand it up somewhere else? Across Microsoft's post, announcement, and product page, the word "leave" never appears. Palantir's fifteen steps are more thorough and no more forthcoming: every one executes on its own control layer, including the step arguing institutional knowhow must sit outside the model. It never asks that same question of the layer underneath it. Declining to train on your data is a promise about the vendor's behavior, one you are asked to trust. Walking out with your loop is a property of your architecture, one you can check. Model choice is the easy half. The substrate the loop lives in, the memory and evals and judgment, is where the lock now sits. Microsoft opened the model layer and left the memory layer proprietary. The lock moved down the stack, and it works the same way one level lower, for the individual whose judgment feeds the firm's loop.

Real company brains, quick toys, and the do-it-yourself ceiling

The gap between a weekend chatbot over your documents and a company brain you can run a business on is wider than a demo shows. The demo hides the parts that only fail in production.

The common build is retrieval-augmented generation (RAG): embed documents, fetch the passages most similar to a question, let the model answer. It breaks in predictable ways. The store keeps serving facts that are no longer true. A vector database does not inherit the permissions of the systems it ingested, so a well-phrased question surfaces a document the asker was never meant to see. And similarity is not correctness, so when no good match exists it returns the nearest thing and the model answers anyway. The hand-rolled version, a fleet of agents over Markdown files, hits the same wall from the other side: concurrent writers corrupt the file, an unbounded file blows the context window, and there is no permission model or provenance at all.

The funded tier is racing to fix this: Sentra and Glean, and a memory-infrastructure layer beneath them (Mem0, Zep, Cognee, Letta). What separates a real brain from a toy is five capabilities, and the field is still weak on the hardest one.

Conflict resolution is the tell. On a public benchmark of agent memory (MemoryAgentBench, arXiv:2507.05257), the best systems reach at most 6% accuracy on multi-hop fact consolidation, the case where the right answer depends on correctly superseding an old one. Similarity search cannot do it, which is the clearest sign this market is still early.

Those five make a brain safe to use. There is a sixth the landscape barely mentions, and it decides whether adopting one deepens the lock-in or breaks it: portability by construction - the memory, its gates, and its receipts export in an open format, so the company can audit it line by line and, if it chooses, leave. Regulation is pushing on all six at once.

The regulation arriving in the EU, India, and Australia

The three regions where many of these customers operate are moving the same way at different speeds. The hard AI-specific mandates are being paced, but the substance is arriving anyway: transparency rules, event logging, human oversight, and data-subject rights, with a data-sovereignty current under all three.

European Union

The AI Act is in force and phasing in: the mid-2026 "Digital Omnibus" moved the high-risk duties (Article 12 automatic event logging, Article 26 human oversight) to 2 December 2027, while GDPR Article 22 already constrains solely-automated decisions and the EUDI Wallet arrives 6 December 2026.

India

The DPDP Act's Rules are operative and its substantive duties bind from around 14 May 2027: itemised consent, data-principal rights, 72-hour breach reports, and, for significant data fiduciaries, annual impact assessment and "algorithmic due diligence" on the software that touches personal data.

Australia

Mandatory AI guardrails were set aside in favour of existing law, but from 10 December 2026 an organisation must disclose in its privacy policy where a computer program makes decisions that significantly affect a person.

Read across the three regimes, the same obligations recur, on a runway that runs to the end of 2027.

The direction is common to all three: logging and auditability as a baseline, human oversight, data-subject rights, transparency about automated decisions, and a steady data-sovereignty pressure underneath (the EU Data Boundary and CLOUD Act debate, India's localisation, Australia's hosting certification).

The window is open, and open is the point

Two clocks are running. The regulation is arriving on a schedule, and the standards it will lean on (digital-identity wallets, authorization grants, AI-Act conformity) are being written now. Systems adopted in this window shape what "compliant memory" comes to mean.

The closed path pours the firm's compounded judgment into a platform it cannot fully audit or cleanly leave. The open path treats memory as infrastructure: a protocol with a reference implementation you can read, run on your own fabric, and export from. Adopt it early, while the conformance profile and vocabulary are still forming, and you help decide them.

Where a delivery partner fits

This fits a partner building software for customers whose data lives in the EU, India, or Australia, where residency and auditability are turning from preferences into procurement requirements. The differentiator, governance, provenance, and the ability to run in place, is what those buyers now ask for.

• Build governed company brains where a no-data-movement Bridge and a per-disclosure receipt turn a compliance obligation into a selling point.
• Differentiate on control: a company brain the customer can audit line by line and export in an open format, set against the closed platforms above.
• Serve the mid-market the incumbents skip. Palantir and Microsoft sell the same sovereignty thesis as a heavy, top-down stack for the largest institutions - Palantir's own playbook doubles as a lead-gen brochure for its Foundry and AIP products; the proportionate, local, partner-delivered version is a market they do not serve.
• Help shape the conformance profile and vocabulary from real deployments across three regulatory regimes, worth more to the standard than any internal design.

None of this depends on the parts still in build. The governed substrate, lens-scoped retrieval, and signed receipts run today, enough to prototype against real data and find the sharp edges early. The question for a first working session: which customer and use case would put the most instructive pressure on a governed memory layer first?

Five questions to ask your customers

Demand for governed memory is easy to test: it surfaces as stalled projects and unanswerable audit questions. Run these five on any existing customer. Two or more uncomfortable answers is a qualified conversation.

1. Do you have AI or copilot projects stalled at security or legal review? maps to · governanceThe commonest blocker is access control the AI layer cannot honor. Gates that travel with each fact are the unblock.
2. When your AI answers from company data, can you show where the answer came from? maps to · provenanceIf the answer is no, every AI output is an unattributable liability under the logging and traceability rules above.
3. Can you prove that who sees what through your AI matches your actual access model? maps to · receiptsA signed record per disclosure is the difference between asserting compliance and demonstrating it.
4. If you changed vendors tomorrow, what happens to everything your AI has learned? maps to · portabilityA loop you cannot carry out is one you rent. The export format decides whether the asset is theirs or the platform's.
5. Who on your side owns the December 2026, May 2027, and December 2027 obligations? maps to · regulationIf nobody owns the dates in section 04, the compliance-as-byproduct story lands hardest here.

Then bring us the most instructive yes: andrew@lexenne.com. Protocol and product detail: lexenne.com/slf · lexenne.com/remember.