Your AI forgets you overnight.

I teach a LinkedIn Learning course on AI Agents and Agentic Memory, and the question I’m getting most right now isn’t “how do I use AI” but “how do I know if the agent my company deployed is reliable?” Agent memory is what lets an AI agent remember what happened earlier in a task instead of starting over blind every time you give it a new instruction. This newsletter gives you the four tests I teach in that course, so you can evaluate what you’ve been handed and make the agents you already have work better starting today.

Move 01 | Test if it holds the thread | 2 min

This is for any multi-step task you’re thinking about delegating. (A Finance team I know found out the hard way when their agent lost context at invoice 47 of 200. Everything after that was wrong, and nobody caught it until reconciliation.)

Before we continue, summarize every decision we've made in this conversation so far, in order. Then pick up from step 3.

If it summarizes accurately, working memory is intact. If it misses steps or contradicts itself, don’t give it multi-step work until that’s resolved.

Move 02 | Build its institutional knowledge | set it up once

Long-term memory is what stops you from onboarding your agent every Monday. If your Marketing team’s agent has to be re-briefed on brand guidelines and client sign-off process every single session, that’s not delegation. That’s a very organized copy-paste.

Here's how my team handles exception requests: [describe your actual process]. Save this. Every time I ask you to review an exception, use this as your baseline — don't ask me again.

Use this pattern for anything you’d explain to a new hire on day one: your team’s approval process, formatting preferences, recurring decisions that always go the same way. Works in Claude Projects or ChatGPT Custom Instructions. Set it up once.

Move 03 | Teach it the format you use | 5 min

This is the difference between output you can send and output you have to rewrite. Your HR team’s agent routing leave requests shouldn’t be using a default approval chain when your engineering team and marketing team have completely different processes. Give it the real example.

Here's an example of the status update my manager actually wants to see: [paste a past one]. From now on, every status update you draft for me should match this structure exactly.

Works for anything format-specific: reports, emails, proposals, meeting notes, exec summaries. Paste the real example, not a description of it. The difference in output quality is immediate.

Move 04 | Stress-test before you trust it with live work | 5 min

If your company has deployed an agent that’s supposed to react to live inputs (monitoring a project board, processing incoming requests, flagging when a dependency shifts), here’s how to find out if it’s handling that correctly.

I'm going to send you three updates over the next five minutes, out of order. Don't respond until you have all three. Then synthesize them into one decision.

If it responds after the first update instead of waiting, its sensory processing isn’t queuing correctly. That’s a meaningful gap to know about before you trust it with anything time-sensitive.

For your IT team | the one question that matters

The four moves above are what you can do right now, on your own, with the tools you already have access to. This last one is for when your company is rolling out AI agents at scale and you want to understand whether they’re going to hold up.

The question isn’t “what agent are we using.” It’s: where does the agent’s memory live, and is it converged?

If the answer is a vector database bolted onto a relational database with a sync layer between them, you have fragmented memory. The agent reconciles data instead of doing your work, and you’ll feel it in reliability before anyone publicly admits there’s a problem.

A converged engine, one that handles vector, relational, JSON, and graph data together with transactional guarantees, is what makes an agent reliable across users and systems and not just on one person’s laptop. You don’t need to understand the architecture to ask the question. Asking it will register.

This week: run Move 01. Paste it into the agent you use most and see what it gives back. Two minutes. You’ll know immediately what you’re working with.

Come tell me on LinkedIn what it said. Specifically what happened when it tried to summarize its own prior decisions. I read every message.

VP @ Oracle | LinkedIn Top Voice for AI | LinkedIn Learning Instructor | Views are my own

AI For You is a weekly newsletter for corporate women (and yes, the guys reading this too, welcome) navigating careers in the AI era. Real tools, real context, no tech background required.