Skip to main content

Your AI agent's memory needs an owner and an expiration date

Workday and Google shipped agents with persistent memory. Bornet wants an owner and expiration date on each memory. Nobody says what happens when staff leave.

Your AI agent's memory needs an owner and an expiration date

Ricardo Argüello

Ricardo Argüello
Ricardo Argüello

CEO & Founder

Business Strategy 4 min read

In mid-October, Workday and Google both launched agents that remember between sessions. On October 16, Pascal Bornet wrote in his newsletter that this creates a new corporate asset, and that someone has to maintain it.

I agree, and I would push it one step further. An agent’s memory is a company asset with an owner and an expiration date, like an archive box with a shred-by label. Right now almost none of them have a label.

What the announcements say, and what they leave out

Workday launched its Autopilot Agents on October 13. In the Workday press release, recruiter preferences such as interview timing or format “can be saved to the agent’s memory.” The Talent Acquisition Agent blog post adds that interview feedback, notes and transcripts “become durable memory.” Every action is logged, including what information the agent used and who approved it.

Google’s Gemini agent announcement, dated October 8, lists four kinds of memory: session, semantic, procedural and episodic, the last one being “everything it has done before.” Each agent gets its own identity, “governed like an employee,” and every action lands in an audit trail attributed to the agent.

So identity, permissions and action logs are there. What I did not find, and I searched all three texts for it, is any mention of how long that memory is kept, when it expires, or who corrects it. It may be in the technical documentation. It is not in the announcements.

Three kinds of memory, three rules

Bornet sorts memory by risk, and his split holds up as written. Convenience memory, meaning formatting, preferences and routines, can persist with light controls. Context memory, meaning customer history, project assumptions and past decisions, needs a source, an owner and a review or expiration date. Authority memory, meaning permissions, prices, contracts, compliance status and entitlements, should not exist at all. The agent checks the system of record before it acts.

His working rule: if a remembered fact can change money, access, compliance or a consequential decision, verify it first. Then test the agent with stale, conflicting and poisoned memories and watch how it recovers.

I would add one sheet of paper. One row per kind of memory, five columns: who can read it, who audits it, when it is reviewed, when it is deleted, and what happens when its owner leaves. Most companies could write the first four today. The fifth is the one almost nobody has written down.

When the person who trained it leaves

This part is my inference. It is not in Bornet and not in the announcements. Say the recruiting agent learned one recruiter’s preferences, and she resigns. Whose memory is that now? Does the next recruiter inherit it, does it get archived, does it get deleted? A departing employee hands back a laptop, badges and mailbox access. Does your offboarding checklist have a line for what the agent remembers about them?

Google’s design makes the question sharper. Coworker agents get their own email address, their own persistent storage, and see only the context you or your team share with them. Part of what the agent knows came from specific people. If those people leave and the context stays, either someone decided to keep it or nobody decided anything and it stayed.

We have covered the two halves of this before. Your AI plateaus if it doesn’t dream between sessions was about consolidating and pruning memory. Altman’s 6 Months: Who Owns Your AI’s Delete Button? was about deletion when the vendor controls retention. This post is about what sits inside your own walls.

What I would do this week

Pick one agent that already remembers things. List what it stores and put each item in one of the three classes. For anything in the authority class, confirm the agent looks it up in the system of record instead of trusting its recollection. For context, name an owner with a surname and put a date in a calendar.

Then run Bornet’s test. Give it an old price and see whether it uses it.

When we run a discovery for AI Maestro, asking what agents remember and who answers for it goes on the same map as processes and permissions. If you want the one-page template, write to me and I will send it.

Frequently Asked Questions

Persistent memory in an AI agent is what the agent keeps between sessions: customer history, project decisions, preferences and unfinished work. A company should govern it because it can go stale, conflict with newer facts or be poisoned, and the agent will act on it with nobody reviewing what it holds.

Pascal Bornet argues that authority should not live in an AI agent's memory: permissions, prices, contracts, compliance status and entitlements. Those get checked against the system of record right before the agent acts, because they change and a stale memory can cost money or open access it should not.

It depends on the company's agent memory policy. If every memory has a named owner, the departure triggers a review: reassign it, keep it with a new source, or delete it. Without an owner, the agent keeps acting on the preferences and assumptions of someone who is gone.

Feed the AI agent stale, conflicting and poisoned memories on purpose and watch whether it checks the system of record before acting. Bornet recommends this test. An agent that uses the old price without looking it up has failed, and it is better to learn that before a customer does.

agent memory AI governance persistent agents Workday Google Gemini Pascal Bornet AI Maestro

Related Articles

Your AI plateaus if it doesn't dream between sessions
Business Strategy
· 9 min read

Your AI plateaus if it doesn't dream between sessions

Anthropic shipped dreaming in agents on May 6. The difference between AI that compounds and AI that plateaus is architectural, not model-driven.

dreaming agent memory agentic architecture
Your AI Wants to Touch Payroll. Kubernetes Knows How.
Business Strategy
· 7 min read

Your AI Wants to Touch Payroll. Kubernetes Knows How.

The engineer who built Azure Kubernetes Service is now Workday's CTO. Container governance turns out to be the playbook for governing AI agents.

AI agents Kubernetes governance
Gemini 4 Argon Gained 14 Points at Box. Run Your Own Test
Business Strategy
· 4 min read

Gemini 4 Argon Gained 14 Points at Box. Run Your Own Test

Box measured Gemini 4 Argon 14 points ahead of 3.1 Pro on its own tasks. How a CTO repeats that test on company documents before switching providers.

Gemini 4 Argon Google Gemini Box AI
The Cheap Model Trap: How AI Providers Capture Ecosystems
Business Strategy
· 8 min read

The Cheap Model Trap: How AI Providers Capture Ecosystems

Google at $0.25/M tokens, OpenAI at $0.05/M. Not charity, it's platform capture applied to AI. What the pricing war means for your B2B independence.

AI strategy vendor lock-in AI costs