# The Factory System - Radexus > Radexus builds institutional memory for revenue. Not a tool your team logs into, but a system underneath the tools that accumulates your company's commercial judgment and applies it to the next decision. *Source: https://radexus.com/factory-system/* --- The Factory System # A company that learns, not twenty people who each got faster. This is what we build, in plain terms. The full architecture is technical and we are happy to walk your CTO through it, but nobody buys an architecture, so here is what it does. **Radexus builds institutional memory for revenue.** Not a tool your team logs into, but a system underneath the tools that accumulates your company's commercial judgment and applies it to the next decision. ## Four properties, and all four have to be true at once. **It remembers** Every quote, visit and decision becomes a case that does not leave when the person does. **It forms judgment** Patterns across hundreds of cases that no individual would live long enough to notice. **It expires** Knowledge carries a confidence that decays, because a system that never forgets is eventually confidently wrong. **It is governed** Nothing enters a system of record without a named approver. Nothing promotes below its gate. Start with the thing every owner already knows. A rep with twenty years leaves, and the company forgets twenty years. Not the records, which are still in the CRM. The judgment: which customers pay late no matter what they say, which discount shape means the deal was already lost, which plant manager decides and which one only signs, why that account went quiet in 2019 and what it took to bring them back. None of that was ever written down, because there was no shape to write it into. So it walked out of the building, and the replacement spends three years rediscovering a third of it. > Individual productivity does not compound into institutional capability. This is why the current wave of AI adoption disappoints at the company level even when individuals like it. Twenty people each twenty percent more productive is twenty better work products and an organisation that has learned nothing. Nothing was retained. Nothing accumulated. The next quarter starts from the same place, with the same institutional ignorance, and the only durable change is a subscription line. You have seen this exact failure before. A CRM rollout records activity faithfully for four years and still cannot tell you why deals die, because recording is not learning. A system that stores every interaction and forms no view of them is a filing cabinet with a search box. ### What memory actually requires Three things, and all three have to be true at once. Any one of them alone produces something that demos well and fails in year two. #### One. It remembers what happened, with the context intact. The unit of memory is a case, not a document. A case is a bounded episode with its circumstances attached: an enquiry, a site visit, a quote and its revisions, who asked for what, what was offered, who decided, and what happened afterwards. It is stitched from your systems and from the things that never reach a system, like a voice note after a plant visit or the reason a quote was pulled. A case is anchored to the resolved entity, so the four spellings of one customer across four systems collapse into a single history. That resolution is the unglamorous work underneath everything else, and it is where most projects quietly fail. #### Two. It forms judgment, not just answers. Retrieval finds the nearest document. Judgment is what emerges across hundreds of cases: that quotes to this customer type stall at week three, that this discount shape precedes a loss more often than a win, that in this territory two site visits before a quote is worth eleven points of conversion, that an AMC lapse plus a competitor visit is a replacement cycle beginning. Patterns are proposed by the system, reviewed by a person, and either promoted with a definition or rejected. A promoted pattern is a written claim with its evidence attached, not a black box. It can be argued with, which is the only way a claim ever becomes something a CFO will act on. #### Three. It expires. This is the part almost nobody builds, and it is the part that decides whether the system is still trusted in year three. Every piece of knowledge carries a confidence that decays with age and drops sharply when contradicting evidence arrives. Prices move. A competitor exits. A plant closes. A policy changes. A system that never forgets becomes confidently wrong, and the first time it is confidently wrong in front of your leadership, the deployment is finished, whatever its accuracy elsewhere. So knowledge that has aged past its half-life is marked stale, held out of decisions, and either re-evidenced or dropped. **Governed learning**Nothing enters your systems of record without a named approver. Nothing promotes below its evaluation gate. Lowering a threshold is a versioned decision with an owner, not a setting someone changes to make a release pass. Learning that cannot be governed is learning nobody will let near a P&L. ### What this is not It is not a chatbot with your documents behind it. It is not a seat your team logs into. It is not a framework you adopt, configure and then staff. And it is not something that improves by pooling your data with anyone else's. **The method improves. Your data never moves.** What we carry from one engagement to the next is how to build memory well: the patterns, the resolution logic, the evaluation harness. Your cases, your judgment and your numbers stay inside your tenancy, are never pooled, and are never used to train anything. ### Why it compounds The graph is built once. The first play is expensive because it pays for the resolution, the definitions and the lineage. Every play after it reads a memory that already exists, which is why the second play in an engagement typically costs a third of the first and the fifth costs less again. The same holds over time. A system with six months of cases makes better proposals than one with six weeks, and it does so without anyone doing anything differently. That is the definition of an asset, as opposed to a subscription. ## Four words we use, and four we avoid. ### What we say - 01 **Institutional memory**Knowledge that belongs to the company rather than to the person who happens to hold it this year. - 02 **Compounding judgment**Every case makes the next decision better, without anyone changing how they work. - 03 **Knowledge that expires**Confidence decays. Stale claims are held out of decisions until they are re-evidenced. - 04 **Governed learning**Named approvers, evaluation gates, versioned thresholds. Learning your auditors can live with. ### What we do not - **AI-powered**It describes the ingredient, not the outcome. Every vendor in your inbox says it, which means it says nothing. - **Copilot**A copilot helps one person fly one plane. Nothing accumulates, which is the entire problem we exist to solve. - **Agents**Whether a play uses an agent loop or a scheduled job is an engineering decision we make inside your build. It is not a product. - **RAG**Retrieval finds the nearest document. It is a technique in the stack, not a capability worth naming to a CFO. **None of those four words survives a CFO asking what it actually does.** Every competitor owns them, which is precisely why they are worthless. A system that never forgets is a system that will eventually be confidently wrong. Which is why expiry is a first-class feature rather than a caveat. Half-lives are set per knowledge type: a price is short-lived, a plant location is long-lived, a decision-maker's preference sits in between. Contradicting evidence collapses confidence immediately rather than averaging against it. The result is a system that says "I knew that six months ago and I no longer do" instead of asserting something that used to be true. That sentence is the difference between a deployment that survives three years and one that gets quietly abandoned in month fourteen. [Start a discovery](https://radexus.com/start/)