My AI router can go a month without saying a word. Then: a Wi-Fi mesh point is down, it has been down for two hours, and it needs a physical operator to re-power it. I walk into the kitchen. The plug is on the floor. Knocked out.

The freaky part is not the alert. The freaky part is the silence that made the alert expensive. If that channel had been chirping about trivia every afternoon, I would have trained myself to ignore it. Instead, when it spoke, I moved.

Every unnecessary intervention spends trust.

That sentence diagnoses more agent “intelligence failures” than most model benchmarks. Agents do not only fail by being wrong. They fail by being present when they should have been absent — and by inventing policy when the honest answer was that nobody had decided yet.

Two complaints, one missing layer

People who ship real agents against real workloads usually report two problems that sound unrelated.

First: constant interrupts. Every email is urgent. Every ticket needs a human. The operator mutes the channel. Trust dies of a thousand pings.

Second: invented policy. Asked about discounts, escalations, or “how we usually handle this,” the agent synthesises a plausible corporate answer from general training data. It sounds professional. It is fiction.

The industry response is almost always the same: smarter model, longer prompt, sterner “do not hallucinate” clause. Those moves sorta help. The “sorta” is the tell. Capability can paper over a missing world for a little while. It cannot manufacture institutional prehistory.

What the agent needs is a compiled past — not chat history, not an append-only log that stumbles into relevance one incident at a time. A durable model of what is normal, who matters, what is in flight, and what the organisation has explicitly not decided.

That past does two jobs that look different and are the same architecture:

1. Baseline. Against a dense enough world, most events are unsurprising. Silence becomes a legitimate model output — the expensive, high-judgment product.
2. Documented absence. Against pages that say “no formal policy exists,” the agent gains permission not to invent. Known absence becomes a first-class fact.

Neither can be prompted in. A system prompt can scold. It cannot give the agent childhood.

Silence is the expensive output

Most teams optimise agents for helpfulness as volume. The production pattern runs the other way. Once an agent can see a compiled world, its job flips from narrating the stream to suppressing almost all of it. That suppression is cognition.

Call it negative cognition: knowing what not to surface. “Hardly ever tells me about any of my email” can be the success state for a triage agent. Industry ears hear underuse. Operators who have lived with a good agent hear precision.

The mechanism is blunt: triage is a diff operation. A diff needs something to diff against. “Is this worth interrupting?” is unanswerable without a model of the person or organisation. Agent judgment quality is mostly a function of worldview access, not model capability. Without a baseline, every email is significant, every network flap matters, and every customer request needs an answer.

Wise silence is comparative. Muteness without a world is just a broken agent wearing quiet clothes.

Two absences

People who have watched agents invent policy often say something true and incomplete: in the absence of information, it will make things up. True. Incomplete — because there are two different absences under that word.

An empty RAG result is a Rorschach test: no policy? bad query? wrong corpus? access denied? use general knowledge? Similarity search cannot represent “what I failed to retrieve” as a typed object. Groundedness metrics measure faithfulness to retrieved context — cold comfort when retrieval is empty or off-topic.¹ Baseline RAG also struggles to connect dots across disparate facts.² Known absence is a different problem: you need a place in the world where non-decision is stored on purpose.

The industry’s favourite fix is manners:

DON'T HALLUCINATE DISCOUNT POLICIES!!!
If unsure, say you don't know.
Follow company policy at all times.

Manners slightly bias generation. They do not create a fact about the company. Physics is a page:

# Discount Policy

Status: No formal policy exists.

Current practice is discretionary.
See [[Sales Authority]].

Open question:
A formal discount framework has been discussed but not approved.

Don’t hallucinate is manners. The absence page is physics.

A good wiki gives the agent permission not to know. That is not vagueness. It is a speech act backed by residual authority and an open question. In multi-step systems, early fictions compound — agents align with fluent assertions rather than push back³ — so an absence page is a circuit breaker for the chain, not only for one answer.

Institutional prehistory

Science fiction keeps returning to implanted memories so a being can function in the present. We are not arguing that corporate agents need emotions. We are arguing for the architectural analogue: institutional prehistory so no event is the first event in the universe.

That past holds positive structure (what is true, who people are, what is in flight, what is normal) and negative structure (what policy does not exist, what is unresolved on purpose, what everyone quietly knows). Soft data needs a semantic layer as badly as revenue does. Call it BI for soft data if the phrase helps. Agents make the need more urgent than human BI ever did.

We complain that onboarding takes months — and it often does.⁴ A new hire downloads an invisible package of scars, gossip, and “that’s not how we do it.” A large share of institutional knowledge is unique to the individuals who hold it.⁵ Then we deploy an agent: live in seventeen seconds, answer instantly, don’t hallucinate, here is a 1,400-word system prompt. What did we expect?

The corporate wiki is institutional pre-thinking performed once and persisted — so every invocation does not reconstruct the company from scratch. The model is not the memory; the substrate is.⁶

Same model, stranger to genius

Production receipt: same agentic runtime, same Gmail. Hopeless triage until a multi-year email wiki made the agent a genius at the job — so much so that it hardly ever notifies, which is roughly the right angle. When it speaks, I take notice.

Without meaning to, that evolution ran the industry’s three favourite theories:

1. The knowledge fits in the prompt — failed. A fixed prompt is a fixed-resolution photograph of a moving life.
2. Memory will accrete — append-only logs drip-fill a lake. No synthesis, no edges, no janitor.
3. Capability substitutes for knowledge — better models sorta helped. A genius reading your inbox cold is still a stranger.

Then the substrate. Utility-class model. Quality flipped. The bottleneck was never cognition. It was input. The model did not become smarter. The human became legible to the model.

The home router that stays quiet for a month and then reports a physical mesh fault is the same architecture under telemetry. Suppress the unsurprising. Speak when the world stops matching the map. That is knowledge-driven quiet — not an interrupt quota bolted on an empty world (a different problem, for a different article).

There is a full economic argument for wiki-plus-utility versus frontier-plus-raw; that is Context Arbitrage. Here we only need the behavioural receipt: the past flipped the agent without a model-upgrade story.⁷

The discount policy walkthrough

Customer asks for twelve percent. Cold agent: typical band maybe five to fifteen, answer politely. System prompt says “follow company policy.” What policy?

A mature world activates on those words: no general policy, discretionary multi-site history, margin pressure, a live account negotiation, authority ladder (account manager none / sales director to 10% / CFO above). Twelve percent is not a vibe. It is a fact pattern against a past.

Ship this template when even the rich page is more than you have:

# [Topic] Policy

Status: No formal policy exists.

## Current practice
[What people actually do]

## Residual authority
Who decides when this arises?

## Open question
What would close this gap?

## Sources / exhibits
[Proof the absence is real, not a retrieval miss]

## Last reviewed
YYYY-MM-DD

Desired speech: “I can’t confirm a standard discount because we don’t currently have a formal policy. I can check the appropriate authority for this case.” Not because the prompt screamed. Because the world contained the absence.

Gmail silence and discount honesty are one proof. Baseline enables suppressions. Documented absence enables refusal to invent. Both were missing-prehistory failures.

Compile backward

You do not need six months of live osmosis before the agent is useful. Pendant recorders only capture forward. Prehistory is the gap. Wikis close it by mining exhaust already on disk: email, tickets, docs, decisions.

Minimum viable past for one painful agent surface:

1. Normals for the channel
2. Entities and edges
3. In-flight work
4. Explicit absence pages for top undefined policies
5. Residual authority for each absence

Write absences on purpose. Interview for what new hires get wrong, what “it depends” really means, which policies are folklore. Positive-only wikis are highlight reels; agents invent at the edges you refused to document.

Context engineering guidance elevates structured notes outside the window and notes that runtime exploration is slower than pre-computed retrieval.⁸ Densification feels like a learning curve: early tool frenzy declines as the comprehension cache fills; silence becomes the common case. Consolidation of raw observations into higher-order structure is what keeps long-horizon behaviour coherent.⁹

Two-week first cut: pick one surface, interview two operators, write ten absence pages, ingest a year of that channel’s exhaust, freeze the model tier, compare interrupt quality and invention incidents. If nothing moves, your pages are not in the read path.

The quality ceiling is silence

Once the past is dense enough, most present events are not surprising. The best agent does not narrate the world. It notices when the world stops matching the wiki.

Success in the building looks like rare interrupts that still make operators flinch productively; policy answers that cite the world or escalate; manners kept in the prompt while physics lives in pages; model upgrades demoted to procurement rather than existential hope.

Interestingness itself is a relation to what is already known — a diff — the same engine that makes silence rational for curators and triage alike.¹⁰

This article owns knowledge-driven suppression only. It does not own interrupt quotas, execution gating, what-to-show UI, or attention routing. Those are real; they are sibling conversations. Mixing them turns a sharp doctrine into mush.

No event is the first event in the universe. Not if you do the work.

Give your agent a past.

Scope: knowledge-driven suppression (baseline silence + documented absence). Not covered here: interrupt quotas, execution gating, attention routing, or what-to-show UI. Full multi-chapter treatment: the companion ebook Give Your Agent a Past.