What arrives in full, and what arrives as a name

Retrieval has two jobs. Choosing the right knowledge is the famous one. Deciding how much room each chosen entry deserves is the one that determines whether the good answer actually fits.

Pathrule used to answer the second question with one word: everything that ranked went out as full content, in rank order, until the room ran out. The entry ranked eighth got exactly the treatment the entry ranked first got, and the entry ranked ninth got nothing at all, even when it was relevant. Relevance decided the order and nothing else.

Now relevance decides the form.

Three forms, one ranking

• The closest few arrive as full content. The assistant reads them and acts, with no separate fetch and no round trip.
• Everything else relevant arrives as a name, an id and one line. Around sixty characters. Enough for the assistant to recognise it and pull the body when it turns out to matter.
• Skills arrive by name only, however well they rank. More on that below.

The important property is the one at the bottom of the list: nothing relevant is thrown away for want of room. Discarding a correct answer because it landed in an inconvenient position is the one outcome that was never acceptable, and the cheap form is what makes it avoidable.

The default keeps the top three bodies, and that number came from measurement rather than taste. Keeping only two would have saved about a tenth of the context by giving up the third-best body on every single prompt. Keeping three preserved it and still cost less than the old behaviour of sending every ranked entry whole. Correctness set the default; the cheaper setting stayed available for anyone who wants it.

Why a skill is never pushed

A memory is a fact. A rule is a constraint. Both are things an assistant should simply know while it works, so both can arrive as content.

A skill is a procedure. Putting a procedure into the context of a turn that did not ask for one does not inform the assistant, it instructs it, and an assistant that has been handed a checklist tends to follow the checklist. So skills are offered by name and the assistant reaches for one when the work is actually the work that skill describes. See writing skills for what a skill is meant to be.

Rules earn their room too

Rules used to bypass the relevance test on the way in. The effect was measurable and it was not subtle: on a run of real prompts, rules were taking the overwhelming majority of the available room while most of those prompts delivered no memories at all. A constraint about database migrations was crowding out the memory that would have answered a question about a button.

Rules now pass through the same relevance test as everything else. The exception is deliberate: a rule marked strict keeps its body unconditionally, because a hard constraint whose violation causes a regression is not something to ration.

The demotion is what makes relevance observable

There is a second reason the cheap form matters, and it is about learning rather than cost.

When every ranked entry was pushed as a body, the layer never found out whether any of them were wanted. A body that arrives unasked produces no evidence either way. An offered name is different: the assistant either pulls it or does not, and a pull is a deliberate act by the party that knows what the work needed.

That pull is the first honest relevance label this system can collect, and the payout for a pulled name is a body slot next time rather than a nudge in a ranking nobody can see. It fits the wider loop described in how retrieval finds the right knowledge: the layer measures what it delivered rather than asking the assistant to report back.

The bar has to hold in every language

The relevance bar is a single threshold, and it turned out to be quietly language dependent. The same memory scores one way against a prompt written in English and lower against a faithful translation of that prompt, because meaning-based matching does not produce identical numbers across languages for the same idea. A bar tuned against English prompts therefore cuts off the other languages without anyone noticing, and the symptom is not an error. It is a team that writes its prompts in Turkish getting less context than a team that writes the same prompts in English.

So the bar is set where it holds for all of them, and the behaviour is verified rather than assumed. Across seven languages, the memory that should have won ranked first in every one of the thirty five case and language pairs tested, and a precisely worded prompt found its entry every time in all seven.

A prompt that asks two things gets knowledge for both

Ask one question and meaning-based matching works well. Ask two in the same message and it has a structural problem: the request becomes a single average of two different ideas, and an average of two things is often close to neither. The entry that answers your first question and the entry that answers your second can both fall short of a bar that either would have cleared on its own.

Pathrule now reads a multi-part request as its parts. Each clause is matched separately, the clauses are matched at the same time rather than one after another, and the results are merged by best match per entry. Two-part questions went from finding knowledge for roughly two thirds of the parts to finding it for all of them in one language and nearly all in another, while single-part questions were unaffected and the time it takes did not move, because the clauses are matched concurrently.

The room available grows with the question rather than with the workspace. Asking two things costs a little more than asking one. Having ten thousand memories instead of a hundred costs the same.

What to read next

• How retrieval finds the right knowledge for the selection this delivery model sits on top of.
• How hooks work for how the selected slice reaches the assistant before its first tool call.
• How Pathrule proves hard claims for the contracts behind the relevance bar and the recorded outcomes when nothing is delivered.