Meet Morrow Works.

Their knowledge is scattered across documents.

A handbook. Meeting notes. A helpdesk copy. Each holds part of the story.

You’re helping this team make that information usable—for colleagues, and for the AI assistants they use.

We will follow one ordinary question. No previous workshop or technical reading is needed.

Then a member asks a simple question.

“How much notice makes cancelling my room booking free?”

The room is in France. The booking was created on 4 May 2027. We are checking the general rule as a member, using information available on 8 May.

These details matter. A rule for another country, product or booking date may not apply.

You search the filenames. Nothing comes back.

A search for “cancellation” finds none of these five filenames. The relevant decision is saved as notes-04-18.md.

Nothing was deleted. Looking in the wrong place is not the same as having no information.

A filename search can still be useful when you already know the name of a file.

Search inside the files, and the picture changes.

Now the decision appears. It says 7 days. But the original terms, a later handbook and a helpdesk copy also appear—and they say 14.

Finding passages is called retrieval. It has found the statements. It has not decided which one applies.

The lookup here really runs on the local text. The documents and policy are fictional.

Three copies do not make three decisions.

The handbook copied the original terms. The helpdesk copied the handbook. Those three 14-day statements share one origin.

The 7-day statement comes from an approved policy-board change. A newer-looking filename does not overrule an approved decision.

We need the relationships between documents—not a vote between matching numbers.

Then you check what the change actually covers.

The board approved the change on 18 April. It applies to France room bookings created from 1 May. The collection received it on 3 May.

Our booking was created on 4 May, and we are answering as of 8 May. It is inside that scope. The supported general threshold is 7 days.

An earlier France booking or a Germany booking would need a different check.

An assistant needs the reasons, not just the number.

For one answer, an assistant receives a selected bundle of information. That bundle is its context.

Here, we supply the rule for deciding authority and scope, plus the applicable decision. A stored source that never reaches this bundle cannot support this answer.

Keeping a file was only the beginning.

We had to find it, understand what it was connected to, check when it applied, and supply the relevant evidence.

And tomorrow, when a rule changes, those relationships and interpretations must stay up to date. That is the memory-engineering problem—not simply storing more text.

This example began with a known question. The wider task also includes exploring unfamiliar collections and discovering questions you did not know to ask.

One answer is not the whole picture.

The member has their answer. Now the team wants to understand what happens when bookings change—not just repeat the cancellation rule.

An operations note offers a starting point: rooms sometimes sit unused after cancellations, while other requests arrive late. Could those released slots be useful to someone else?

We are still reading as a member, with information available on 8 May 2027. These are qualitative notes, not measured results.

One note can belong in several useful views.

Put this note in an “operations” folder and you have a place to store it. But you may not think to open that folder when studying booking changes.

Give the same source several ways in: room capacity, the booking lifecycle, and preparation work. These are views onto one note—not three new copies.

Start with useful, revisable categories. You do not need to invent a perfect filing system before exploring.

Before connecting things, check what they are.

Elsewhere in the collection, “Atlas” appears twice. One is a room in France. The other is a recording kit in Germany.

A shared name might put them in the same search result. It must not turn them into one thing—or carry the room’s rules over to the kit.

A useful grouping answers “What is this related to?” Identity answers a different question: “Which thing is this?”

A promising connection is a question, not a conclusion.

Return to the unused-room note. A separate draft proposes a two-day cancellation threshold for a future France-room pilot. It also asks for rebooking and preparation costs to be compared.

Putting the draft beside the observations gives the team something useful to investigate. It does not prove that the pilot will fill empty rooms—and it does not make the draft an approved rule.

Keep the link and its uncertainty together. The observations do not establish the effect of changing the cancellation window.

A useful overview keeps those differences visible.

The team can now see a small picture of the problem: unused rooms have been noted; a pilot has been proposed; actual rebooking remains unmeasured in these notes.

Compress that into “shorter notice improves occupancy” and the uncertainty disappears. Keep observed information, proposals and missing evidence separate, with a route back to the sources.

An overview is another representation to inspect and correct—not a replacement for the underlying material.

Now the collection helps you ask the next question.

“Are released room slots actually booked again?” That question was not the member’s original request. It emerged by following the relationships between the notes.

The next investigation needs booking and rebooking evidence, then a way to evaluate the proposed change. This collection cannot establish a rate or a causal effect. That is an evidence gap—not permission to invent an answer.

You have moved from finding a known answer to discovering what would be useful to learn next. A clear gap is a useful result.

The rule changed before the collection caught up.

You already met the seven-day rule. Rewind to 3 May 2027: the collection has just received those board minutes. The decision was approved on 18 April and became effective for eligible bookings on 1 May.

Those are three different events: a decision, its applicability, and its arrival in the collection. Receiving the document late does not move the policy’s effective date.

We are looking back from 8 May as a member. This replay does not invent a new decision after the previous chapter.

“New” does not mean “replace everything.”

The decision changes the general threshold for France room bookings created on or after 1 May: seven notice days. France room bookings created from 1 January 2027 but before 1 May keep their original fourteen-day threshold.

Germany rooms and equipment rentals are outside this amendment. The collection must attach the change to the right part of the knowledge—not overwrite every cancellation rule.

France-room examples here concern ATLAS-ROOM-FR, viewed as a member on 8 May. These are general terms, not checks for individual waivers.

The source is up to date. The summary is not.

Imagine the team saved a short interpretation on 2 May: “France room bookings require fourteen notice days.” It reflects the older terms available to the collection, but misses the amendment that was already effective.

Now the minutes have arrived. That saved interpretation does not repair itself merely because another file was added. A colleague or assistant could still be given the stale summary.

The received decision remains S03.

Correct the interpretation. Keep its trail.

A revised interpretation separates the two cases: fourteen days for France room bookings created from 1 January 2027 but before 1 May, seven days for those created on or after 1 May. It points back to the original terms, the amendment, and the applicability rules.

The earlier snapshot can remain in the audit history, clearly marked incomplete for eligible bookings from 1 May. It must no longer be supplied as an unqualified current answer.

Retaining an old belief explains what the system represented. It does not prove that belief was correct at the time.

Two different questions can both be about the past.

“What applies to a France room booking created on 29 April?” asks about the booking’s terms. Looking back on 8 May, the general threshold is still fourteen days: that booking is outside the amendment.

“What information had the collection received by 2 May?” asks a different question. It had not yet received S03. That information gap does not mean the approved change was not effective.

Both questions concern ATLAS-ROOM-FR and a member’s general-policy view. The first uses knowledge available on 8 May; the second reconstructs the collection’s 2 May evidence boundary.

Some contradictions cannot be fixed by picking a winner.

Back on 8 May, the team finds another problem—this time for ATLAS-KIT-DE, the recording kit in Germany. For returns from 6 May, one logistics lead says before 17:00 UTC; another says before 18:00 UTC.

Both instructions were recorded on 4 May, both leads have the same authority, and neither instruction supersedes the other. Paragraph order is not a tie-breaker.

Keep both instructions, mark the return deadline unresolved, and request an authoritative logistics decision. A confident-looking combined answer would conceal the conflict.

This is a member-readable handover question, not a room cancellation rule. No resolution is supplied by the fictional sources.

The member knows the booking number. That is not a key.

Back at the member desk on 8 May 2027, someone asks about DEMO-R17, the France room booking created on 29 April. We can explain its general cancellation terms. Can we also promise what this particular booking will cost?

Not from the member’s view. Morrow Works reserves booking-waiver records for finance. Knowing the booking number does not grant access, and the member’s answer must not confirm or deny whether a hidden waiver exists.

The useful response is narrower: explain the general policy and recommend an authorized booking-specific check. A source being stored somewhere is not permission to use it for everyone.

Scene scope: member · ATLAS-ROOM-FR · DEMO-R17 · booking created 29 April 2027 · knowledge on 8 May.

The right decision is found. The assistant still misses it.

Now take a different, hypothetical booking: the Atlas room in France, created on 4 May 2027. The member asks on 8 May for the general free-cancellation notice threshold—not a final fee for a particular cancellation.

Imagine that the lookup has found both the original terms and the approved amendment. But the application sends only the original fourteen-day passage to the assistant. The amendment stays outside the packet it receives.

Here, context means the information supplied for this particular answer. Finding the right evidence and putting it into that context are separate steps. The failure can happen between them.

The hypothetical 4 May booking is within S03’s France-room scope; the member may read both sources.

Send the scope with the evidence.

The team repairs the illustrated packet. It now carries the question’s entity, country, product, booking date, reader role and knowledge date, together with the relevant applicability instructions and policy passages.

For this France room booking created on 4 May, the approved amendment selects seven notice days. The original terms and scoped change remain traceable, rather than becoming a floating number with no explanation.

This is not a request to send the whole collection. Select what is permitted and relevant, and keep the boundaries needed to interpret it. More text is not a substitute for the right evidence.

Same hypothetical case: member · ATLAS-ROOM-FR · created 4 May 2027 · knowledge on 8 May.

An included source can still be used badly.

Even with that repaired packet, an assistant could turn seven days into a universal rule. Germany rooms, equipment and earlier France bookings would then be swept into a change that does not cover them.

A supported answer for this example is: ‘For this France room booking created on 4 May 2027, the general free-cancellation threshold is at least seven notice days.’ It is not a calculation of an individual cancellation fee or a disclosure about private exceptions.

Now the whole chain is visible: stored, found, applicable and permitted, included in context, correctly used. Inspect each transition. Passing one step does not certify the next.

S03 leaves Germany rooms and equipment unchanged; S02 and S03 bound earlier France bookings. Missing case details require clarification, not guessed defaults.

A few days later, the question is which door to use.

Move forward to 11 May 2027. The same member is arriving at ATLAS-ROOM-FR. A temporary instruction, received on 9 May, says to use the courtyard entrance for arrivals from 10 May through 12 May during maintenance.

For an arrival on 11 May, that instruction is current. Its end date matters as much as its content: it expires at 00:00 UTC on 13 May.

The instruction has also reached an extracted fact, an overview and an answer cache. Which of these copies depend on the source?

New time and task: member · ATLAS-ROOM-FR · arrival and knowledge on 11 May 2027. S10 was not available in the earlier 8 May scenes.

On 13 May, yesterday’s instruction is not today’s answer.

At 00:00 UTC on 13 May, the courtyard instruction expires. For an arrival that day, it must no longer answer the current-entry question. The source does not tell us which entrance applies afterward; do not invent a replacement.

At that time, the retained instruction may still explain an arrival on 11 May. Historical usefulness and present applicability are different questions.

Expiry is not deletion. The records rule allows this short historical interval, then requires the temporary instruction and its owned derived copies to be purged at 00:00 UTC on 15 May.

Member · ATLAS-ROOM-FR · knowledge on 13 May 2027. Contrast a current 13 May arrival with a historical 11 May arrival. The allowed history interval is specific to this fictional rule, not a universal retention policy.

On 15 May, forgetting has to reach the copies.

At 00:00 UTC on 15 May, the modeled obligation reaches four owned places: the operational source store, extracted facts, derived overview and answer cache. Removing only the original would leave its content elsewhere.

The required end state keeps a content-free receipt with the source ID and deletion time—not the entrance instruction. From these owned stores, the modeled system can no longer answer which gate was specified for the 11 May arrival.

Commercial policy decisions and waiver history are outside this deletion rule.

Member · ATLAS-ROOM-FR · knowledge at or after 00:00 UTC on 15 May 2027 · historical arrival on 11 May. The deletion rule covers stores owned by the team.