For as long as companies have kept a knowledge base, it has had one kind of writer: a person, writing slowly, with something to lose. Every page cost someone an afternoon, so the library grew at the pace of considered work. That constraint was invisible, and it was doing more work than anyone realised. It was the quality filter.
That constraint is gone. Agents write. They summarise what they touch, document what they build, and file what they conclude, at a volume no team of people could ever produce. Much of it is useful. But point that firehose at the library your company relies on and the library stops being one. A knowledge base with machine writers pollutes by default.
One document at a time, it looks harmless. An agent that documents everything it touches produces a procedure for tasks nobody asked to have documented. Work in progress gets filed mid-flight, half finished but formatted like the finished thing. A project is quietly stopped and its plans stay behind, confident and findable. Two agents pick up the same question on the same day and file two answers that disagree. None of these is dramatic on its own. Together they are how a decade of clutter arrives in a quarter.
The damage lands twice. People can no longer tell which page a colleague wrote and which one a model produced, so they stop trusting the library, which was the point of having one. And agents read what agents wrote: the next run inherits contradictions the system itself manufactured, and acts on them at full confidence. In an agentic company, pollution is not an inconvenience. It is compounding error.
The reflex is to make the knowledge base read-only for machines and call it solved. That is half right. The knowledge base does need protecting. But output that has nowhere to land does not disappear. It scatters into chat logs, ticket comments and download folders, the useful part of it becomes unfindable within a week, and people start re-asking agents questions that were answered last Tuesday. You need both things at once: a knowledge base machines cannot pollute, and a place where everything they produce can land.
The way through is to stop treating company knowledge as one store with one level of trust. Give it four tiers, each with its own rule for what enters, what leaves, and how much authority it carries.
The knowledge base is the top tier: fixed and human-approved. Small on purpose, current on purpose, a name behind every page. Nothing enters it by being written. Things enter it by being accepted: a person reviews, takes ownership, and the page becomes canon. The machine proposes what to remember. A person decides what becomes memory.
Company memory is the working tier, and it is where everything an agent produces lands by default. Drafts, summaries, analyses, work in progress, and the ambiguous material every system handles worst: two drafts that disagree, a claim that could not be verified, a summary that contradicts a standing procedure. All of it searchable, timestamped, attributed to the run that made it, and none of it carrying authority. Ambiguity sits here visibly, owned by someone and with a shelf life, instead of hardening into a second truth. A fresh output that contradicts the canon deserves attention, because one of the two is wrong: either the knowledge base is stale and gets updated, or the output is off and the correction itself is worth keeping.
The archive is for what faded or was superseded: not worth working from, not yet worth destroying. It sits outside the working set of both people and agents, so it can no longer be confidently wrong in anyone's hands. You go there on purpose: a dispute, a post-mortem, an old decision that needs revisiting.
The bin is for what you keep only because a retention rule says you must: financial records, personnel files, whatever your regulators name. Sealed from daily search and from agents alike, and destroyed when the clock runs out. Keeping everything forever is not the safe option. Under European data-minimisation rules, storage without a purpose is a liability with a shelf life of its own.
The movement between tiers is the whole design. Up is a decision: a person accepts something into canon, one page at a time. Down is disuse: what nothing has relied on lately fades from memory into the archive, and out of the archive when the retention schedule says so. No tier fills by accident, and only one tier is allowed to speak with authority.
The naive way to rank what matters is by how often it is referenced, and it is a trap. Citation count rewards the past. The most-referenced page in a store is often exactly the superseded one: the old pricing model everything once pointed at, the process the whole company used to follow. Rank by raw connections and the stale thing stays on top forever, which is the precise problem you were trying to solve.
The rule that works is older than software. Human memory does not archive by age; it archives by disuse. What gets used stays vivid, and each use resets the clock. Score company memory the same way: every reference counts, discounted by how long ago it was last relied on. A fact the system leaned on yesterday is at full strength, however old it is. One untouched for six months sinks toward the archive on its own, before it gets the chance to be confidently wrong.
The effect is not subtle. In a test graph we score this way, the most-connected decision in the store, eleven references, ranks third. First place goes to a decision with a third as many references, made last week. Raw count loses to recency, which is exactly what you want from a memory.
Two guardrails make this safe. Some knowledge is pinned and never fades: the incident runbook is not less true because the quarter was quiet, and standing policy does not expire through lack of drama. And nothing is destroyed by the machine alone. Decay moves things down; only a person signs off on the last step. The machine proposes what to forget. A person decides what is forgotten.
A knowledge base that stays small and true is the difference between agents acting on your business as it is and agents acting on it as it was. It is also the part a competitor cannot buy. Anyone can rent the same model. A clean, current canon of how your business actually works has to be built, and it compounds: every decision made against it is a little better, and every correction that flows back through the gate makes the next one better still.
In Pulse this is mechanics rather than policy. Agent output lands in working storage by default, promotion into the knowledge base passes a gate with a named owner, and the record keeps who accepted what, on what basis, when. Decay and retention run on schedules a person signed off. One rule everywhere: writing is cheap, becoming canon is gated.