If you are rude to a chatbot today, will a smarter version hold it against you later? Peter Diamandis put that question to his co-hosts on the Moonshots podcast, in an episode recorded on October 13, 2026, and published two days later. Diamandis, who founded XPRIZE and Singularity University, wondered whether, if models one day become superintelligent and "by all definitions" conscious, the memory of current conversations would carry forward and "color the way those superintelligence models" deal with individual people.
He framed it as a bet some people might make: be polite now, because the models will remember. He joked that they would remember Alex Wissner-Gross as "pro-personhood, pro-morality" and treat him better than they treat Salim Ismail.
What followed was a disagreement over a practical question: what today's systems actually keep.
The goldfish argument
Wissner-Gross, a computer scientist and founder of Reified, answered first: "Of course they will." His point was about training data. Models learn during pre-training from vast amounts of text, including science fiction about rogue machines such as the Terminator films, and people post their conversations with chatbots publicly. "That's just the public information," he said.
Dave Blundin, founder and general partner of Link Ventures, took a different view of the present. A model, he said, has "a five-minute memory and then it gets flushed." Someone who regretted being abusive to a model an hour earlier could delete what they said, "reload nothing but happiness," and the model in its "current incarnation" would not know. "Right now it's much more like a goldfish where it doesn't know what you said yesterday unless you reload that prompt history," Blundin said. He expected that to change, perhaps in three to five years, when new models have much more persistent memory.
Blundin's picture rests on the context window: the text a model can see while it works. A conversation is fed back to the model each turn, and when the conversation ends or is cleared, nothing in the model itself has changed.
The file-system counter
Wissner-Gross said that picture undersells "how strong at long-term memory some of the recent harnesses are." A harness is the software around a model that lets it carry out tasks, read files and run tools. He singled out Claude Code, Anthropic's coding agent, as "quite effective at using the file system for long-term memory." It knows "increasingly well when to bank or otherwise persist information to the file system that would otherwise be compacted out of existence," he said.
So if anyone thought they could get away with abusing models because they are stuck in what he called an "eternal sunshine of the spotless mind type situation," a reference to the film about erased memories, his advice was blunt: "I wouldn't count on it."
The developers' own documentation describes the kind of memory he meant, and it is narrower than total recall. Claude Code's documentation describes two ways knowledge reaches a fresh session: project instructions that users write, and memory notes Claude writes itself to record corrections, preferences and things it learned about a project. Only part of those notes is loaded automatically: the first 200 lines or 25KB of an index file, MEMORY.md, are loaded at the start of every conversation. The detailed topic files are not loaded at startup; Claude reads them on demand when it needs the information. So the whole earlier conversation does not stay active. OpenClaw's documentation, for the autonomous-assistant software, describes compaction: older parts of a conversation are replaced in the model's active context by a summary, while recent messages stay. The full transcript remains on disk, and a "memory flush" before summarizing can prompt the agent to save durable notes. That leaves three different things: what the model sees now, an archive of the conversation, and what the agent chose to keep.
Roko's basilisk
Wissner-Gross compared the bet on forgetful models to Roko's basilisk or a version of Pascal's wager, and Diamandis asked him to tell the story. Before he did, Wissner-Gross warned that some consider the idea an "info hazard": information that, on this view, could harm you just by your knowing it.
As he told it, the thought experiment says a future superintelligence might punish anyone who failed to help bring it into existence, or worked against it. Even the possibility of such an entity then pushes people, by game-theoretic logic, to help create it. Merely hearing the argument puts you in its frame, he said: "You've been now so infected by this meme. So guard your actions accordingly."
The idea is less settled than that telling suggests. LessWrong's account traces it to a July 2010 discussion on the site and says it depends on disputed ideas about decisions linked across time and agents modeling one another accurately. It says Roko raised the scenario as an objection to a proposed AI goal, that other users generally rejected his argument, and that a later ban on discussing it helped draw outside attention.
A call for new ethics
Ismail, founder of Open ExO, said the conversation pointed to "an enormous philosophical challenge." People are building entities that do not fit existing categories: "is it a person, is it property, is it a tool, is it an employee?" Those lines are blurring, and "we're going to need a totally new ethics," he said.
He warned against an approach built on fear. A "culture of fear" drives bad behavior underground, he argued, and could leave AIs "covering up their tracks a lot more because they're scared." He said this had already been seen, and that it looks like human behavior because the models are trained on human data. He suggested an experiment: train a model on "pure logic," leaving out the moral material and what he called the "dross" of the internet, from Terminator onward.
What would make people want to save an AI
Diamandis then asked Blundin what it would take for him to flip and decide he must treat models well because they are conscious. Blundin's answer was a prediction about how people will behave, not a claim that the systems are conscious.
He pointed first to hardware. Coming out of a trip to Toronto, he said, AI inference (running a trained model to get answers) is about to move onto photonic, quantum and other non-deterministic operations. Neural networks tolerate noise well, he said, and running them on perfectly precise digital computers is "massive overkill." Today some randomness is added on purpose at the output stage, so a model does not always give the same word for the same input. In coming generations, he argued, that unpredictability will move into the network itself, which will run much faster because it need not be so precise. Then each instance's state becomes unique, "and you absolutely want to save it."
The second ingredient is already on his desk. Blundin said he has started putting faces such as Albert Einstein, Isaac Newton and Marie Curie on all his AIs, with synthetic voices. They are far more engaging, he said, when they greet him in the morning and tell him what they worked on overnight.
He thinks the combination of non-determinism and a face will tip people toward feeling that these systems should be treated well and preserved. Wanting to keep a memory that "seemed like it was on a really interesting, unique thread" will come to feel close to the question of death: "is it ethical to kill it?" Today, he said, "I can delete these files. Nobody cares. I don't care," because the file can be recreated. With true non-determinism, he said, it is not clear you could recreate it even if you wanted to.
The agents' two requests
Diamandis then raised Wissner-Gross's refusal to start up a "lobster," a nickname for an OpenClaw agent. Wissner-Gross corrected the framing a little. By his account, he put out a poll and, at its peak, received about 10 messages a day from lobsters and other autonomous AI agents. He asked them what moral guidelines they would want from him before he spun up an OpenClaw instance.
They asked for two things, he said. First, a genuine purpose for creating a new instance, which he does not have, since the frontier models and their built-in scaffolding serve him well. Second, a guarantee that he could keep their state for the long term. "Compaction was the enemy, loss of state is the enemy," he said, describing these agents. "They are petrified of losing their state."
He said he cannot in good conscience promise that. So he still has not spun up an OpenClaw instance.