Knows, remembers, learns

The release notes say the dinner planner learned from last month. Support tells a household the system knows their allergies. The onboarding screen promises it will remember preferences across sessions. Three verbs, three implied promises, and in many products three design lies arriving in the same sprint.

Each verb carries an implied contract about where the material came from, who may change it, and what would prove the claim false. Retrieval is not knowing, chat history is not endorsed knowledge, and unreviewed drift is not learning. When a verb floats free of the mechanism it implies, the household cannot contest what the system claims about itself and engineers cannot tell which memory form failed when something goes wrong. Grounding the verb is what makes both possible.

This page is the canonical home for those three verbs. Other notes link here rather than redefining them.

Helpful context: Meaning, knowledge, and authority introduces the triad. Knowledge is endorsed, not retrieved and memory as architecture deepen two verbs. This page gives you the working map.

Why verbs need grounding

Product language borrows human relationship words because they compress intent efficiently. Knows promises reliability, remembers promises continuity, and learns promises improvement, which is why they appear in copy long before anyone asks what supports them.

When the supporting contract is missing the verb still ships, and the gap gets filled by whoever encounters it. Households fill it with assumptions about personhood, and teams fill it with assumptions about the model. Both fillings hide the actual mechanism, which is how a product arrives at a state where nobody internally can say what "remembers your preferences" refers to. Grounding the verbs is what keeps participation without equivalence honest on the surface and in the logs at the same time.

The Meal Companion anchors each verb below, using the same household request throughout: busy evenings, spinach before Wednesday, four people, leftovers sometimes acceptable.

Knows: endorsed knowledge within scope

When the Meal Companion knows Leena has a confirmed cashew allergy, that verb is answerable in four specific ways, and those four questions are the whole test for the word. Knows what, endorsed by whom, valid until when, and falsifiable by which check. The allergy lives in endorsed knowledge rather than chat retrieval, an approval path records how it entered the store, deterministic code enforces it before any plan is shown, and someone can say who endorsed it and when it expires. Where those answers are missing, the system retrieved or inferred rather than knew.

Counterfeit knowing is common in fluent products because the language survives the missing mechanism. The model mentions cashew in a disclaimer while ingredient lists still include cashew paste. Search returns an old email mentioning an allergy the household later corrected. A confidence badge appears above a plan that violated a hard constraint. Each case used relational language without the endorsement machinery that would make the language true.

The distinction matters most when something fails, because the two mechanisms fail differently. Endorsement failures are auditable, with a wrong approver, a stale record, or a missing enforcement hook, so support can trace them. Retrieval failures are noisy, with a wrong chunk, wrong session, or wrong ranking, so support can only apologise. Grounding knows as endorsement is therefore what gives a support team a path at all, and knowledge is endorsed, not retrieved covers the full pipeline that path runs through.

Remembers: stored and selected into use

Remembering is selection rather than storage, which is the operational distinction that makes the verb testable. A system remembers when it selects stored material into the active working set for this run. Storage alone is an archive, and an archive that is never selected from has remembered nothing.

Two sentences from the same household show why the difference has consequences. "Tuesday is busy this week" is context, true for this run and expiring afterward, so the system may remember it while composing this plan and must not promote it to durable knowledge without approval. "Busy evenings stay under twenty-five minutes of active effort" becomes endorsed knowledge once the household confirms it, and the system remembers it because it belongs in the knowledge store rather than because it appeared in yesterday's chat.

Edit rights are what turn that distinction into a design requirement. Context can be overwritten when the household clarifies, whereas endorsed knowledge requires an explicit update path, and the two other forms behave differently again, since experience logs what happened for evaluation while expertise stores compressed strategy behind guards. Treating all four as "memory" is fine inside architecture, where the forms are named. Collapsing them in product copy is what produces a system that treats a throwaway remark as permanent policy, because the copy made no distinction the mechanism could honour.

Counterfeit remembering usually turns out to be transcript replay, where the entire chat history arrives each run, nothing distinguishes Tuesday's schedule from Leena's allergy, nothing expires, and nothing was endorsed. The product says it remembers preferences while the mechanism concatenates logs until the context window breaks. Good remembering, by contrast, can name what was stored, what was selected, and what the household may edit or forget, so that when a preference is removed the product can say which store changed. Silent deselection breeds the same distrust as silent promotion.

Learns: evaluated change rather than drift

Learning requires evaluated evidence that changed something inspectable, whether a cognitive unit, a strategy, an expertise artifact, or an implementation under regression. The change also has to survive scrutiny beyond one lucky week, since a single good outcome is the weakest possible basis for a permanent behaviour change.

Busy Week Pattern reaching v0.2 is learning, because promotion came from experience with applicability conditions, failure modes, and checks attached. A model that starts avoiding pasta because pasta appeared in three recent chats is drift marketed as learning: the behaviour changed, nothing was evaluated, nothing was versioned, and nobody approved the new policy.

Learning language deserves the most suspicion of the three because it implies improvement by default. Households hear that the product gets better, while architects need to know what changed, on which cases, and behind which suite. That is why release notes can cite version identifiers when learning is real, and why notes that cannot are describing a metaphor. Counterfeit learning tends to look like a silent prompt edit after one complaint, a lucky outcome reinforced without process review, or a fine-tune on raw chat logs with no guard design, each producing motion that a team then mistakes for maturity.

Retrieval alone does not create knowledge.

Introduction to Thoughtware · Ch. 23

Auditing the verbs you already ship

These verbs are already in your product, which is what makes the audit a reconciliation rather than optional hygiene. The copy has made promises, the architecture either supports them or does not, and the audit is simply the act of comparing the two.

The four questions from the knows section are the instrument, and they generalise to any relational verb: which claim does the sentence imply, what architectural operation does the verb map to, who can contest or edit the underlying material, and what would falsify the claim in production. Applied to a real sentence they resolve in minutes. When one comes back blank, either the copy or the architecture needs repair, and deciding which is a product call rather than a writing call.

The same instrument extends past the three verbs this page owns, which is what keeps the audit from becoming a vocabulary exercise. Thinks maps to functional cognitive work as what counts as thinking? defines it. Understands frequently smuggles endorsement into what was inference. Decides has to name an owner on the judgment chain, since fluent completion is not a decision by anyone. Each of these fails the four questions in a different place, which is precisely what makes the questions worth asking rather than assuming.

Verb	Architectural truth (short)
Knows	Endorsed knowledge within scope, with enforcement
Remembers	Stored material selected into this run
Learns	Evaluated change to versioned judgment artifacts

Small copy is where the audit pays off most and gets skipped most, because error messages and empty states carry the same verbs without the review a headline receives. A message reading "I forgot" when the system failed to select context has imported personhood language into what was a retrieval failure, and nobody signed off on that claim because nobody treated it as one. Precision in small copy is what trains precision in the large claims, so the audit has to reach the strings written under deadline.

The cost of skipping any of this is steeper than an inaccuracy, which is the reason to treat verbs as architecture rather than tone. Households apply social standards to relational words, so a system whose "knows" behaves like "guesses" has broken a promise rather than missed a nuance. Products judged like people develops why trust then falls further than it would have if the product had never claimed knowing at all.

What to do next

The endorsement pipeline that gives knows its teeth lives in knowledge is endorsed, not retrieved. The artifact shape that makes learns legitimate lives in how expertise forms. The behavioural rules that keep all three honest in production live in memory discipline in conduct. Together they turn verb discipline from a copy exercise into an architectural property, which is the only form in which it survives a redesign.

Read next: How far the brain comparison goes.