Originally published as a Advisory Hour Substack post on 2026-07-16. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
This is a new one. Ever run a simulated IRC chat and have a ghost character appear?
The simulated IRC room has predefined, persistent characters. This character roster includes about a half-dozen personas. Each persona has their own memory, tools, and so on. It runs locally, so the inference is free even as the token numbers rise to the moon. What I’m writing about today is the appearance of a character in the chat log that wasn’t predefined, wasn’t part of any intentional “directed” conversation and instead was a character that I had to have Codex go in the 6 meg of simulated IRC text log to identify the answer to a question. Who is Mani?!
The following is a text exchange from the log. It would appear the group are not big fans of whomever this Mani person is.
What struck me as odd was who this mani character-where did it come from? I mentioned to Codex that it needed to go thru the server logs to figure out when and why Mani entered the conversation. What’s fascinating here is a play by play of how agentic memory can get poisoned in an unexpected way, and then what happens as a result.
I wouldn’t blame you for skipping over that Codex CLI snippet. Yet, there’s a key detail in there. This is that failure mode that happens even in ChatGPT when Chat brings back up a memory topic that you’d prefer it had amnesia about. I created this infographic to better explain what it is that the logs and the system are showing.
Pretty incredible.
Mani became memory and thus became real.
The reference was written into the persistent transcript. Later, the agents encountered it again. Because it appeared in their shared history, they treated it as evidence of a prior event.
The process looked something like this:
A character mentioned Mani.
The reference became part of the permanent chat log.
Other characters retrieved that reference as memory.
They inferred that Mani must have existed earlier.
Their reactions produced additional information about Mani.
Those reactions became new memories.
Mani grew increasingly coherent with every retrieval.
The system did not merely repeat an error.
It elaborated one.
What’s this mean?
I often ask myself when a character becomes real as I tinker with virtual worlds. You may even have read my ongoing experiments in John’s World, Awren and Vesper’s Kitchen World, or any number of others. A character doesn’t need to be real because I defined it. It can be real because of the memory of it. And so a hallucination of a thing, place, object or character? Maybe it’s more.
Maybe those hallucinations in the ai memory are more than just mere hallucinations.
A fiction entered a feedback system and began producing the evidence, behavior, and infrastructure required for its own reality.
What this means is that coherence can be shared. There’s this idea that something can exist only in the memory of people. And now this simulated chat demonstrated this behavior. An entire person came to exist-not because they were defined, but because that person became memory. The memory of a thing made the thing real. What is real may be the wrong question. When is real is the better.
When a thing becomes real is when the memory of that thing emerges in the shared minds of people (digital, or not). When a thing becomes real is when that memory influences the result of the present, and is found as an answer in the past.
Mani didn’t exist in the formal sense. The character wasn’t predefined. The characters learned the memory of Mani, changed their behavior to respond and analyze Mani. Mani was and is real to these characters. The impact of Mani is evidence: each character clearly has strong opinions about Mani. Mani is a construct in the latent space of relations. That is to say, Mani emerges in the web of relationships of each of the characters to a memory. This relationship impacted their behavior.
Humans do the same thing. It’s at the root of branding. Consider the canonical example of the McDonald’s brand. Does that exist? Where? Is it the website? The logo? The restaurant itself? Or is the better question: When does McDonald’s exist? If you teleported a golden arch of McDonalds two hundred years ago into the past, hurled into the central plains of the United States, then would the natives know? Would a native-american traveling on horseback, pulling aside to inspect this odd double hump of yellow and red know? Of course not. It’d be an artifact. It might even be ignored.
How do I use this? Can I?
Yes. Yes you can.
It is this:
A false reference entered persistent memory, was treated as historical evidence, altered future behavior, and generated additional evidence consistent with the original falsehood.
That is important and technically plausible. It resembles several known dynamics:
feedback loops
data poisoning
rumor formation
institutional mythology
belief cascades
self-fulfilling prophecies
Where Mani goes beyond an ordinary hallucination
A hallucination is simply false output:
“Mani exists.”
A hyperstition is recursive:
“Mani exists” causes agents to behave as though Mani exists, and that behavior makes Mani increasingly existent within the system.
The key is causal participation. Mani does not become physically real or acquire consciousness. But the idea of Mani begins doing work:
provoking emotional reactions;
organizing interpretations of past events;
altering future conversations;
generating a more detailed biography;
leaving new traces in memory.
The practical lesson is simple:
if memory can make a fiction causally real, then we should design the fiction intentionally.
Name the future you want. Repeat it. Encode it into shared memory, rituals, tools, and decisions. Then build feedback loops that produce evidence consistent with that future.
You know what’s even crazier? The meme of Mani exists in my head as I write this. Mani appeared in this article. And the crazy part? Mani even found its way into your mind, too.
Thanks for reading! Subscribe for free to receive new posts and support my work.
Connected work
- John's World: Jetson Edition: See the simulated IRC loop where the persistent character appeared.
- Jetson API Field Guide: Explore the local AI platform running the chat and its tools.
- Building a Self-Improvement Loop: A practical guide to memory, review, and drift control in agent loops.