Every morning I opened the session and started over.
Not the work. The explaining.
Who the client was. What we decided last week. Why that one file is the way it is. The three constraints that are obvious to me and invisible to the tool.
I typed all of it again. Then again the next day.
For months I told myself the model was the problem. A smarter one would remember. A bigger context window would fix it. I was waiting for the tool to catch up to me.
Here is the part I did not want to admit.
The tool was never going to hold my project. I was.
The whole mental model lived in my head. The agent had the last twenty messages and nothing else. Every session it woke up with amnesia, and every session I paid the tax of rebuilding the world before any real work started.
If you build with an AI agent, you know this exact morning. You open it. It forgot. You start explaining. And somewhere around the third paragraph of re-context, you wonder if you are actually behind, or if the tool is.
You are not behind. You are missing tiers.
That was the turn for me. I stopped treating memory as one thing the model either has or does not have. It is not one thing. It is a stack.
There is the memory of THIS conversation. Short, sharp, gone when the window closes.
There is the memory of YOU. Your preferences, your voice, the way you like things done. That should survive every session, and for most people it does not, because it lives in their head instead of anywhere the agent can read.
There is the memory of the PROJECT. The decisions, the constraints, the reasons behind the code. This is the expensive one to lose, because losing it means re-deciding things you already decided.
And there is the memory of what is TRUE right now. The live state. What shipped, what broke, what changed since yesterday.
Most teams use the first one. The conversation. That is it.
The retention you actually want lives in the other three. None of them are automatic. You decide what belongs in each, and you give it a home the agent meets every time it wakes up, instead of a home in your head that you keep performing out loud.
I am going to say the uncomfortable opinion plainly.
The context you re-explain every morning is not the model failing you. It is a design gap in how you set the agent up. The fix is not a smarter model. It is deciding, once, what the agent should always know.
When I finally did that, the mornings changed. The session did not wake up empty. It woke up knowing the client, the constraints, the last decisions, without me performing them again.
The speed came back. The bigger thing was the mental load. I had been carrying the whole project in my own head as the single source of truth, and quietly resenting the tool for not sharing the weight. The moment the weight moved out of my head, the work got lighter in a way I did not expect.
There is a tradeoff hiding in here, and it is worth naming.
Writing things down so the agent remembers is slower in the moment. It feels like overhead. Typing it fresh each time feels faster, because it is faster, today. The cost is invisible because you pay it in small daily slices instead of one visible bill.
I paid that invisible bill for months. Small slice, every morning, never noticed, always there.
So here is the real question, and it is not about tools.
What does your agent actually need to know to be useful to you, and where does that live right now? If the honest answer is in your head, and you retype it, that is the gap. Not the model.
I am still tuning which tier holds what. Some things I thought belonged to the project actually belonged to me. Some things I treated as permanent were live state and went stale fast. Getting the tiers wrong is its own lesson.
But I stopped waiting for the model to remember my project.
I decided to give it a memory instead.
Your turn
How do YOU stop your agent from forgetting your project between sessions?
If this was useful
I work through this in public, the wins and the freezes both, mostly on LinkedIn and YouTube. If the real version of building in the open is useful to you, that is where it lives. Find me on X, GitHub, and the work at next8n.com.
Top comments (0)