Two days ago, I turned ChatGPT’s memory back on.
I wanted to see how far the “improved” experience had come.
Late last night, I turned it off again.
The experience had become weird.
I would ask an agent to help with one problem, and it would pull in some unrelated matter from the past. An old preference. A previous project. A temporary issue we had already solved.
Then it would try to force that old material into the work in front of it.
I kept catching myself thinking:
“Why the hell are you talking about this?”
I was close to building yet another memory function to fix the problem.
Then I stopped and looked at what had actually changed.
The only thing I had done was turn ChatGPT’s memory on.
So I turned it off.
The weirdness stopped.
That pointed to retrieval, not reasoning.
A lot of AI memory works through what people call RAG, short for Retrieval-Augmented Generation.
The name sounds complicated. The mechanism is simple.
Imagine an AI has a library full of old conversations, notes, decisions, and documents.
When you ask it a question, the system searches that library, guesses what may matter, and puts a selection in front of the AI before it replies.
That can be useful.
It can also go badly wrong.
A weak retrieval system does not simply leave an irrelevant note in the library. It promotes that note into the current conversation. The AI sees it as part of the job, then spends energy trying to make it relevant.
More stored context can produce worse current judgment.
This morning, I saw Theo’s video on turning off Claude Code’s memory. He was talking about coding agents, but the underlying point was immediately familiar.
For an AI to work well in a codebase, it needs to know what the codebase is, where things live, and how to find the right thing when it needs it.
The source code is the truth.
The file system is the map.
The agent needs a clear route to the information that matters now.
That is why I built Memory Loom.
Memory Loom gives memory both a proper home and a controlled way back into the work.
Shared Memory holds company facts, decisions, procedures, and routes that more than one agent may need.
Agent Memory holds the continuity of one specific agent: its role, its working relationship with me, what it owns, and the local lessons it has earned.
Those memories are stored as structured, versioned OKF records. Each record has a clear identity. When something changes, we can update the current record while preserving the earlier revision instead of letting two competing versions float around forever.
The agent can call named Memory Loom tools to read a current record, save a correction, revise a decision, or retire something that has become stale.
Storage handles history.
Retrieval handles relevance.
When a new conversation begins, the Context Loader gives the agent a Context Map. It can see what is already loaded, what other records exist, and where to retrieve deeper detail if the task genuinely needs it.
That is a very different experience from asking an AI to carry every old note into every new task.
And it matters enormously for Partner.
Partner is our system for forming agents around a real, ongoing responsibility rather than a one-off prompt.
A Partner has to survive the conversation ending.
It has to reopen the work with the right decisions, the right boundaries, and the right understanding of what changed. It needs enough immediate context to continue responsibly, plus a clear route to deeper evidence when the work calls for it.
For a Partner, memory becomes operational continuity.
On 31 August, I’m hosting a 60-minute community call where I’ll share more about the answer we are building:
- self-improving, autonomous agents;
- agent memory that stays useful as work changes; and
- how Partner can help an AI carry real responsibility inside a business.
Register here:
[Archived link retired.]
I’ll share what worked, what failed, and the lessons I had to learn the hard way.
FELIX