Continuum
All essays
Essay

Context that earns its place

Why decision history can become a durable asset without becoming a trap.

October 2026

Two systems can use the same model and give very different advice. One knows the words in the request. The other understands the circumstances: what the user is trying to protect, which exception mattered last time, and why an apparently sensible option was rejected.

We think that difference will become more valuable as models become more capable and widely available. A capable model supplies intelligence. Context determines whether that intelligence is applied to the right problem.

Our thesis is that a small number of personal and enterprise platforms can become deeply embedded by learning how their users decide. But collecting history is not the same as understanding it. And a platform that traps history may create dependence without creating better decisions. The durable asset is useful, governed decision context, not a larger archive.

What an archive misses

A record might show that a company rejected a supplier. It may not show whether the reason was price, a temporary capacity problem, an unacceptable contractual term or a relationship that management wanted to preserve elsewhere.

Those distinctions determine what should happen next. If capacity was the issue, a later offer might deserve reconsideration. If a contractual restriction remains, the same offer should fail again. Remembering the outcome without its reason can make the next decision worse.

Useful context therefore includes the decision, the evidence available at the time, the constraints that shaped it and the consequences that followed. It also distinguishes explicit rules from inferred preferences, durable priorities from temporary conditions, and information the system may use from actions it may take.

This is not simply a demand for longer context windows. More text gives a model more material to read. It does not ensure that the model finds the relevant fact, notices that it has changed or knows when the evidence is insufficient.

LongMemEval, an academic benchmark presented at ICLR 2025, tests information extraction, reasoning across sessions, temporal reasoning, knowledge updates and abstention. Its significance for this thesis is the separation of those abilities. Remembering something accurately and using it appropriately later are different problems. The benchmark is evidence of a technical challenge, not evidence that any particular company's memory is a moat.

Personal taste and company culture share a structure

For a person, context includes taste, habits, relationships and priorities. A useful system should learn the difference between a stated preference and a choice made under pressure. One inconvenient booking should not become a permanent preference. A preference from a year ago should not silently outweigh a clear instruction today.

For a company, the equivalent includes workflows, approval boundaries, recurring exceptions and the practical rules behind formal policy. Much of what people call culture appears in decisions: which risks deserve escalation, when a customer exception is worthwhile and what quality standard is non-negotiable.

We see a common product problem underneath these different settings. The system needs a current account of what matters, why it matters and who has authority to decide. It must be able to apply that account without assuming that yesterday's answer governs today's circumstances.

The resemblance does not imply one universal product. Enterprise context is divided across teams, contracts and permission boundaries. Personal context may involve intimate information and conflicting priorities within a household. The useful abstraction is shared; the governance and distribution are different.

How history becomes an advantage

A history becomes valuable when it changes the next decision for the better. A system sees an exception, records why it was accepted, observes the outcome and learns which part of the earlier reasoning deserves to be reused.

That improvement need not require continuous retraining of model weights. Better retrieval, updated rules, improved evaluations and clearer representations of the task can all improve performance. We would not confuse a product that updates its knowledge with a model that learns continuously. Those are different mechanisms, and both need testing.

The potential advantage is cumulative. A newcomer can license the same model, but it cannot instantly reconstruct the relationships among past decisions, operating constraints and observed results. It may have the documents and still lack the account of why those documents mattered.

There are limits to this argument. If history is noisy, poorly attributed or rarely relevant, more of it can add cost and confusion. If users must spend hours correcting the system's assumptions, the supposed advantage becomes a maintenance burden. And if a company has no legal right to reuse customer data for shared improvement, it cannot count that reuse as an asset.

Our durability test is straightforward: give a competitor the same model tomorrow. Which measured advantage remains?

The equal-model test

We would want to see the contextual product outperform a strong baseline on the user's actual work. That means comparing like-for-like models and permissions, and separating useful personalization from easier tasks or a better interface.

The improvement might be fewer repeated questions, fewer exceptions mishandled, faster completion or better decisions under changing conditions. It should also include the cost of maintaining the context. A product that saves ten minutes and requires twenty minutes of correction has not passed.

We would test what happens when a fact changes. Does the system retire the old assumption? Can it explain the basis of a recommendation? Can the user correct that basis without rewriting their entire history? Does the system abstain when the record cannot support a confident answer?

A platform earns its place by doing those things consistently. The customer should stay because replacing it would mean accepting worse performance, not because exporting their information is impossible.

The sovereignty tension

The more a system understands a customer, the more useful it can become. The same accumulation can make the customer more exposed if the provider controls the only usable account of their decisions.

That is the tension at the center of contextual AI. We want products that improve through use. We also want customers to retain control over the information and authority that make those products useful.

Sovereignty does not necessarily mean running every model locally, owning every weight or refusing external services. It means having practical control over where information goes, who can use it, what the system may do and which capabilities survive a provider change.

A portable download of raw conversations is better than no export, but it may leave the useful structure behind. Decision reasons, source references, corrections and permission boundaries matter too. Portability should preserve enough meaning for the customer to keep using its knowledge, even if another system must adapt the representation.

We do not expect every product to export its proprietary methods. A vendor can own its evaluation tools and software while the customer controls its records and authorized uses. Distinguishing those assets is more credible than claiming that every improvement belongs exclusively to one side.

Control belongs in the product

A contract can promise that data remains the customer's. The product still needs to make that promise usable. Can the customer inspect an inference, correct it, limit it to a team, revoke access or delete information that should no longer influence a decision?

Authority needs similar precision. Knowing that someone usually approves a transaction does not authorize the system to approve the next one. A history of behavior is evidence about preference, not a grant of power.

NIST's zero-trust architecture describes granular access decisions and least privilege. Applying that principle to contextual AI is our analysis: access to an information source should not silently become permission to use it for every purpose or to act on everything it reveals.

This can improve adoption rather than obstruct it. Customers can entrust a system with more useful context when they can bound its use. A product that requires broad, irreversible permissions asks users to take the greatest risk before it has proved the greatest value.

The strongest objection: incumbents already have the context

Existing platforms own email, documents, calendars, customer records and operating systems. They have a large starting advantage in access and distribution. Better models may make the information they already hold much easier to use. A new contextual platform could be left paying to reconstruct what an incumbent supplies by default.

We take that seriously. An independent product needs to do more than connect more data sources. It needs a materially better account of the work: superior handling of changing facts, better expert judgment, cross-system reach or a governance model customers cannot get elsewhere.

A second objection is that portability weakens the moat. If customers can leave with their context, what remains? Our answer is performance. The product can still own better retrieval, evaluation, updating and execution. Requiring captivity to preserve differentiation suggests that the differentiation is thin.

There may be fewer winners than the number of products advertising memory. The opportunity is in understanding consequential work, not storing every conversation about it.

What we look for

We look for contextual products with an identifiable decision advantage, a credible way to keep facts current and explicit rights over how information is used. We want evidence that outcomes improve with continued use, not an assumption that retention proves quality.

We are skeptical of passive archives marketed as intelligence, opaque preference scores that users cannot correct and products whose usefulness depends on unrestricted data access. We would also reject a supposed learning advantage that relies on rights the company does not have.

This thesis weakens if general platforms can reproduce contextual quality from portable records with little loss, if historical information adds no measurable value beyond the current task, or if maintaining it costs more than it saves.

The lasting asset is not everything the system knows. It is the part of that knowledge that improves the next decision, under the customer's control.

Sources

The technical and security sources below support the underlying distinctions. The durability, sovereignty and market conclusions are Continuum's analysis.