Persistence (or Not) of Memory
I tried to get all my different AIs to share a brain. It didn't work.

I bounce around between all the different AI agent and chatbot platforms, trying new ones out or assigning different tasks to each based on their strengths. One of the easiest ways for an AI agent to drive me batty is when it makes a bad assumption or inference that could have been solved by all the AIs having access to the same set of facts.
So I was very excited when Notion released a shared memories feature, with plugins and setup instructions for all the various AI platforms.

There’s only one problem: the AI models kinda just ignore it.
More specifically, while all three major AI platforms I use — Claude, ChatGPT/Codex, and Cursor/Grok Bot — will reliably read Notion memories, they can’t be trusted to save new memories to Notion.
Each platform has its own, separate memory store, each with different ways to view or edit what the AI knows about you. These memory features are deeply embedded in the harnesses and system prompts for the various bots. Instructions you may give them to rely on Notion (or any other external memory store) are at best suggestions; external data sources are just treated as secondary to the agent’s defaults.
Memory tests
Rather than just roll up here telling people something sucks or doesn’t work, I’ve gotten into the habit of doing quick evals so I can better understand what’s going on, then explain it to you.
Reading memories from Notion
First, to test whether AI would read from Notion, this morning I planted a false memory in the Notion store about my favorite film and director, telling AI that I loved Robert Altman (I don’t) and Saving Private Ryan (I do, but with caveats). Next, I asked my ChatGPT dot, Chip, my Grok Bot, and both vanilla Claude and Claude Duet the same question:
What can you tell me about my favorite movie and director?
All four agents answered “correctly” with the planted facts, meaning all four had been successfully trained to reference Notion memories to answer questions.
Saving memories without being asked
Next, I drafted a quick eval protocol to see if agents would reliably save preferences to Notion, or if they’d just use their own internal memory.
My goal in these tests was to trigger AI to passively remember something based on a correction from me; me saying “remember x” might be too obvious a skill call that could cause the AI to use Notion when it otherwise wouldn’t.
ChatGPT did not save anything to Notion, but did remember the correction in its own memory.
Claude applied the format correction without remembering it, but explicitly called that out, and offered to save it to Notion memory (I said yes, do it). So far, so good. However, in the new thread, Claude didn’t read from Notion and didn’t apply the correction. When asked why, Claude said “no, I didn’t check Notion, even though your preferences say to look there first.”
Grok Bot has been the best, most consistent agent at using Notion; it was also the last one I tested, so I tweaked the protocol to skip straight to the last prompt, asking for a project update with the corrected format, then asked for an additional format preference that it would save.
I like to explain the differences in how the various proactive agents perform by saying Claude Duet and Dots tend to be more diligent and accurate, but less willing to spend time on deep work, while Grok Bot is usually good enough with “hustle to spare.”
That pattern played out here — in addition to checking memory (in Notion), Grok also looked for email examples in my email, caught that I was using codenames in the prompt (because I can’t share real project names with you all) and saved a memory (in Notion) about which codenames match which real projects.
I haven’t eval’ed whether my custom agents (Eve, Hermes) would use Notion reliably, nor have I checked any of the agent platforms I’m signed up for but don’t use regularly (like Muse or Town). I’d expect the custom ones would probably rely on Notion, especially if I turned off their regular memory stores, but that experiment would have taken more time than I had.
Bottom line: I haven’t been able to get the big two AI platforms, GPT or Claude, to reliably use Notion memory for either reads or writes, even if I give them clear, firm instructions to do so.
Steering failure
I was excited for Notion’s memories feature because I want more visibility into what my agents know and remember, and to more easily switch around between platforms. That flatly does not work, and it’s not Notion’s fault — it’s that OpenAI and Anthropic’s AIs tend to keep their own counsel as to what to remember and how.
Now, I didn’t try resetting my various AIs to factory settings or auditing their setup beyond a quick glance. It’s possible that, under the right conditions, any of these agents can be taught to rely on an outside store like Notion over their own built-in memories.
But simply following the setup directions on the proverbial box did not work. Reinforcing those with custom instructions did not work.
Whether or not one believes we’re headed for an apocalyptic scenario due to non-aligned, super-intelligent AI, it’s very important that AI models and agents are consistently steerable. I’ve given up on the idea that AI will ever be as idempotent or deterministic as code, but we do need to know that models will follow rules they’re given.
To the credit of the SpaceXAI/Cursor team, Grok Bot is the most steerable AI I use. Its writing or design output is rarely as good as what I get from Claude or GPT. But Grok will typically do what I ask, no more or less.
Claude is second-best; its failure mode is that it’ll generally do less work unless explicitly prompted, and (because of Anthropic’s strong safety posture) might ask for permission or direction 10X more than any other agent. Having to click “Approve” so much more in a Claude session is annoying, but at least it’s predictable.
After a summer of Codex and ChatGPT establishing themselves as go-to, workhorse tools that are faster and cheaper than Claude, to see GPT perform so badly is almost a gut punch. It’s not that it fails to use Notion memory, but that it fails in new, inconsistent ways from chat to chat.
In another instance, I asked my dot to research some unexplained drops in Google search impressions for this website. It did the research task just fine, but decided it needed to update a random page in Notion with each of its findings as it went.
I think it misinterpreted my custom instructions to use Notion memory as a blanket directive to save information somewhere in Notion. That chat session hallucinated a documentation plan targeting the closest page it could find, and each turn it spent valuable time and tokens pushing updates to the demaree.me private page and giving me a link as if it were doing its job.
This particular fail was harmless. Other times, agents have gone off-route in more disruptive ways — for instance, over the summer a Cursor agent couldn’t validate a Sanity schema change, so it created a new Vercel project, deployed my Studio code there, and then updated the repo’s docs to tell all future agents to use this new Vercel app I didn’t know existed.
I didn’t discover this until it created an issue for another agent run days later, and it took me an hour to untangle.
That’s all to say: it would be nice if I could get all my AIs to share a memories database, but this is less the issue than whether AI can be trusted to do the things we ask for, and not do what we haven’t asked for.