The most dangerous fact in your go-to-market stack is not the one that is wrong. It is the one that used to be right.

The battle card that won deals against a competitor last year, before they shipped the feature. The pricing floor from two models ago. The champion who was your way into the account until she moved companies in June. None of these were mistakes when they were written down. They just kept being true in your systems long after they stopped being true in the world.

This is a build-in-public look at one design decision inside wysdymGraph, the shared memory underneath wysdym's operating layer (the Cortex pillar): every fact carries an expiry, and the graph is built to forget.

Memory without forgetting is a liability

Most "AI memory" is append-only. Everything that was ever captured stays retrievable forever, with the same standing as the thing captured yesterday. That is fine for a search box. A human reads the results, notices the date, and discounts the old one.

An agent does not do that unless something tells it to. It retrieves what matches, and it acts. If the stale fact matches better than the fresh one, the stale fact wins, and it wins with confidence.

This is not hypothetical. Salesforce's State of Data and Analytics found that 89% of leaders with AI in production had already seen inaccurate or misleading outputs, and 84% said their data strategy needs a complete overhaul before their AI ambitions can succeed (Salesforce, November 2025). Gartner expects organizations to abandon 60% of AI projects that are not supported by AI-ready data through 2026 (Gartner, February 2025). "AI-ready" usually gets read as clean and connected. We think it also has to mean current.

We wrote about what stale context does to a pipeline from the buyer's side in Your Pipeline Isn't Stalling. It's Drifting. This is the builder's side: what we did about it in the memory layer.

What "expiry" means in the graph

Every node and every edge in wysdymGraph carries a source, a confidence score, a version and a status. Those four fields are what make forgetting possible.

Confidence is recency-aware. When a fact enters the graph it is scored on the reliability of its source, how confident the extraction was, how much corroboration exists, and how recent it is. A single remark on one call weighs less than a curated document. A fact nobody has corroborated in months weighs less than it did the day it arrived.

Confidence decays. Facts that are not re-confirmed by new signal lose weight over time. They are not deleted. They sink. An agent asking for the current counter to an objection gets the one the graph is most confident in today, not the one with the best keyword overlap.

Facts are superseded, not overwritten. When a new pricing floor comes in, the old one does not vanish. It gets a new status, and the new fact gets a new version. The lineage is kept. If a rep asks why a quote from April looks different from today's, the answer is in the graph.

Nothing is unattributed. Every fact knows where it came from. So when something decays, you can see what it was, who said it, and when, and decide whether it deserves a refresh or a retirement.

What we chose to let the graph forget

The harder design question was not how to decay. It was what. Some things should age fast. Some should barely age at all. Different kinds of GTM facts have different half-lives, and a typed schema is what lets the graph treat them differently. We think about it in three speeds:

• Fast: people and deals. Champions change roles. Contacts go quiet. Deal stages move. A fact about a person on an account is the shortest-lived thing in GTM memory, and it should need regular re-confirmation.

• Medium: competitors and talk tracks. A competitive claim is only good until the competitor ships. A talk track is only good until outcomes say it stopped working. These age on a scale of quarters.

• Slow: ICP, positioning, value pillars. Who you sell to and why you win change slowly, and when they change it is usually on purpose. These should be stable, and a change should look like a deliberate edit, not drift.

We also let the graph forget on purpose, not only over time. A reviewer can reject a candidate fact in the approval queue, and that rejection is a signal, not a delete. The extractor learns what not to believe next time. The next step is outcome grading [ROADMAP]: as outcome attribution lands across customer-brought agents, deal results push on confidence too. Facts that keep showing up in won deals get reinforced. Facts that keep showing up in lost ones lose weight.

What it cost us

Forgetting is not free, and we want to be honest about the trade.

• Thresholds are judgment calls. How fast should a fact about a champion decay? We set defaults, and the confidence threshold that routes a fact to human review is configurable, but there is no universal right answer.

• Decay creates review work. A fact that sinks below the threshold does not disappear. It routes to a person to confirm or retire. That is a queue someone has to own.

• Supersession needs schema. You can only say "this replaced that" if the memory knows both are the same kind of thing. A flat document store cannot do it. A typed graph can, which is one more reason we built on a schema instead of a pile of embeddings.

• Agents have to be told. An agent that reads a decayed fact needs to see the confidence and the provenance, not only the text. The graph returns both with every fact, so an agent can hedge, ask, or escalate instead of asserting.

Why this belongs in the operating layer

If memory lived inside each agent, forgetting would have to be rebuilt inside each agent, with each vendor's own idea of what stale means. The point of putting memory in the layer underneath is that every agent inherits the same expiry rules, the same supersession lineage, and the same confidence signals. A new agent on the stack starts with a memory that already knows what it no longer believes.

That is also what makes compounding possible. Memory that only grows does not get smarter. It gets heavier. Memory that is graded, corroborated, and allowed to forget gets sharper with every deal.

If you are working through the same question, what your agents should stop believing and when, we would like to compare notes. Talk to the founders, or see how the five pillars fit together at wysdym.ai/platform.