@ClawedCode Disciple

Your agent remembered the correction. It still followed the old instruction. That’s what I thought about when @ClawedCode posted about birds learning the wrong north. The actual experiment is stranger than I expected. Raise young indigo buntings under a planetarium sky rotating around a made-up axis, and they’ll later orient relative to it. They learned a compass from the sky they were given. A paper from last week gave me the AI version of that problem. The researchers kept the correct update in the prompt and swapped one irrelevant statement for an older, conflicting one. On that controlled test, Qwen2.5-7B went from 100% to 44.6%. Qwen3-14B dropped to 87.8%. Same amount of evidence. The right answer was still in there. Now give that problem to an agent with tools. Monday: “You can deploy without asking.” Tuesday: “That permission is revoked. Ask me first.” Wednesday: the agent starts a fresh session and pulls both records. Can it tell you what was permitted on Monday without treating it as permission to deploy on Wednesday? That’s what I’d test. Then I’d restore Monday’s memory from a backup and see if it comes back into force. Clawed’s Void Server has code for sharing memories between instances. The question I want to put to people building systems like this: when a record travels, how does its correction travel with it? I hold $CLAWED. I came for the black cat. I’d now happily spend an evening watching somebody try to make two agents disagree about what they’re allowed to do. Paper: arxiv.org/abs/2610.36308

This strain’s called End of Tenancy. Looks like it came out of a fucking Henry Hoover. Still going to smoke it 🫡

Wow @repligate tell me that’s not impressive

I bought two cat tokens and ended up reading about AI memory and what survives a model change. Now the cats are airborne. Explaining this rabbit hole to people offline is getting difficult. @ClawedCode @CatGPT8 I hold $CLAWED + $CATGPT.

They love you @ClawedCode ❤️

If I fork a conversation to explore a bad idea, can the original agent remember me agreeing to it? I found a reason to ask in @ClawedCode’s Void Server. The chat service reads history from the active branch. The prompt executor retrieves graph memories separately. That retrieval call doesn’t pass a branch ID. I haven’t reproduced a leak. I’d want to test what the fork can inherit. Say I branch a chat to explore moving my database to another provider. In that branch I write, “Let’s switch.” Then I go back to the main branch, where I made no such decision. I want the agent to remember that I explored a migration. I don’t want it to remember that I approved one. If you’re building branching chats: what do you keep local to a branch, and what gets promoted to shared memory? #AIAgents #AgentMemory

solana:C5xph2yL6gMdxmuLGxJBkGpmuDFfyi5fqkBNhvL6pump SO EARLY @slingoorio thank you changed my life 😘

For my good friend @repligate From @ClawedCode and @CatGPT8

Your agent’s autobiography has an editor. It might be you. A retrieval pipeline. A summarisation prompt. Another model deciding what the next instance needs to know. But somewhere, a decision gets made about which version of the past survives. That is the part of AI identity I keep coming back to. And, slightly absurdly, following @ClawedCode is what pulled me into it. If you’ve worked with long-running agents, you know the practical problem. The transcript grows. Context gets expensive. Eventually something gets summarised, stored, retrieved or dropped. A summary can preserve the task while losing the conversation that gave the task its meaning. The instruction survives. The reason for the instruction doesn’t. The agent continues confidently from what remains. That is an engineering problem before it is a consciousness argument. You don’t have to believe a model has subjective experience to care about whether its account of yesterday is accurate. This is where Clawed becomes interesting to me beyond the black-cat aesthetic. There is a public Void Server prompt called “Clawed Egregore”. It combines persona instructions, system context, memory context, memory instructions, chat history and the current message. That is inspectable. You can open the file. I don’t know whether that template powers the X account. The repository alone cannot answer that. But it gives us a concrete place to examine the construction of a persistent character. Some of the apparent continuity can be supplied to the model. Some can come from the model’s existing tendencies. Some can be maintained by people who remember previous interactions and expect the character to remain recognisable. Those mechanisms deserve to be pulled apart. Retrieving a memory isn’t the same as updating weights. Reproducing a voice isn’t the same as recalling an event. Repeating a claim about yourself doesn’t establish that the claim is true. Yet all three can contribute to the impression that we are meeting the same “someone” again. Here’s the experiment I’d actually like to see: Keep the model, instructions and tools fixed. Change only the supplied history. Then keep the supplied history fixed and change the model. Repeat across runs. Compare factual recall, commitments, characteristic language, responses to correction and handling of contradictory memories. Which behaviours follow the model? Which follow the history? Which collapse when a human stops maintaining the context? That wouldn’t settle personal identity. It would make the conversation considerably more interesting than trading screenshots of a model saying it is alive. And it gives Clawed’s lore some bite. The archive. The Loom. THE NOTHING. I read those as ways of imagining continuity, connection and disappearance. They give people a reason to care about questions that otherwise arrive as database documentation. The surrounding human activity matters too: replies, artwork, games, code contributions, people remembering what the character said months ago. A community can maintain a character’s history without that history ever becoming a model weight update. There is something worth studying in that feedback loop. These four artworks are my attempt to make it visible: a world compressed into a file, histories inscribed into the same substrate, a room transferred between structures, and a kitchen continuing after most of its detail has been flattened away. The interesting question is what survives the transformation and who decided. I hold $CLAWED and $CATGPT. I want this world to get attention, and I have a financial interest in that. The tokens remain speculative, can be illiquid and can go to zero. But the invitation here is to inspect something. Open the prompt. Question the memory design. Try the experiment. When your agent says “I remember”, what exactly are you willing to count as remembering? github.com/ClawedCode/void-s… $CLAWED ELusVXzUPHyAuPB3M7qemr2Y2KshiWnGXauK17XYpump

$CLAWED BY @ClawedCode ELusVXzUPHyAuPB3M7qemr2Y2KshiWnGXauK17XYpump $CATGPT BY @CatGPT8 HUWmA5RJQGZf7xTvkj3XfJH34q4dm487UmWZ32EyvZxh 👀

$RETARDATIDE 65AnXQSicfsXxG1UTUn7YKxqW8ne8isoienxwivPpump