No Persistent Memory
Every session starts from zero. Your context does not carry, and nothing announces that it didn't.
This page covers one specific way AI gets things wrong at work, and what to do about it.
It runs in order. What goes wrong, why it happens, where you'd notice it on an ordinary day, who takes the blame, roughly what it costs, and the check that catches it. Then one thing to try this week.
The dollar figures are estimates, not measurements. The assumptions behind each one are printed right there, so you can swap in numbers that fit your job. Anything actually measured carries an OBSERVED tag.
What is no persistent memory?
The model is stateless between sessions. Whatever context exists comes from what is in the current window — pasted, retrieved, or injected by a memory feature that is a separate subsystem with its own failure modes.
Where this becomes expensive: corrections do not persist. You catch an error, you correct it, the model acknowledges the correction and produces the right output. Next session, the same error. Not because it ignored you — because the correction existed only inside a window that closed.
Memory features complicate this rather than solving it. They store some things, selected by heuristics you don't control, retrieved when a matching rule fires. So you now have partial continuity, which is worse than none: it teaches you to expect carryover that only sometimes happens.
What do people assume?
That a system used repeatedly accumulates knowledge of your work — that the decisions, corrections, and constraints established last week inform this week's output.
Absent a deliberately engineered memory layer, each session begins with no knowledge of any prior one. The interface implies continuity. The system provides none.
Where does it show up at work?
A pricing analyst spends a session establishing that the model must use net-of-rebate figures, not gross. Correct output. Good session.
Three weeks later, a different analysis. Gross figures. Same tool, same account, no warning. The correction was never institutional — it was conversational, and the conversation ended.
Who carries the downside?
Vendor: none. Executive: decides on the number. Manager: owns the analysis. You: produced it. And you will be asked why you made the same mistake twice, which is a materially worse conversation than making it once.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Recurring analyses / year: 150 Rate reintroducing a previously corrected error: 10% (~15 / year) Rate caught in review: 80% Errors reaching a decision: ~3 / year Cost per error: $5,000 – $40,000
Annualized exposure: ~$15,000 – $120,000
How do you control for it?
Externalize corrections. Every correction becomes a written standing instruction in a reusable Context Brief (Part III, Concept 2) — not a message in a thread.
The test: if the correction lives only in a conversation, it does not exist.
CONTROL COST Standing brief maintenance: 2 hours / month Annual: 24 hours Fully loaded rate: $70 / hour
Annualized control cost: $1,680
What should you do this week?
RECOMMENDATION
Start a standing context document for your function. Every constraint, definition, and correction goes in it. Reuse it at the start of every consequential session.
Then notice what you've actually built: a written specification of how work in your function must be done. That document has value with the AI, without the AI, and during the next onboarding. You have written the thing that survives you — which is the definition of institutional knowledge, and it is a rare thing for an individual contributor to own.
Evidence
ANALYSIS Statelessness between sessions is an architectural property, not a research finding — state as design fact, not citation.
This is one of 24 failure modes. The book gives you all of them — plus the controls that catch each one and a 90-day plan to prove you ran them.
Preorder the Book