Latest / Ctrl AI Profit / Ep. 185 | OpenAI Just Changed What Its AI Remembers About You
Transcript
- 0:20SPEAKER_01: I'm Michael, a small business owner and entrepreneur since 1983, founder of Cadenhead Services and 850 Media.
- 0:27SPEAKER_01: I speak from four decades of real operational experience, not white papers.
- 0:31SPEAKER_01: This is Control AI Profit, and this week OpenAI changed how its AI remembers everything you tell it.
- 0:39SPEAKER_00: OpenAI rolled out Dreaming V3, a new memory architecture for Chat GPT that replaces the old saved memory system with background synthesis.
- 0:49SPEAKER_00: Instead of requiring users to explicitly save information, the system now automatically processes past conversations, documents, and connected apps to build a persistent memory profile for each user.
- 1:02SPEAKER_00: Internal tests report factual recall improved from 41.5% to 82.8%, with a roughly 5x reduction in compute cost.
- 1:12SPEAKER_01: This is a bigger deal than a typical product update.
- 1:15SPEAKER_01: For years, AI memory worked like a notepad.
- 1:18SPEAKER_01: You had to manage yourself.
- 1:20SPEAKER_01: You told the model, remember that I prefer brief responses, or remember that my company uses Salesforce.
- 1:26SPEAKER_01: If you forgot to save it, the AI forgot it.
- 1:29SPEAKER_01: Dreaming V3 is different.
- 1:31SPEAKER_01: It reads your entire conversation history, identifies what matters, and builds a profile automatically.
- 1:36SPEAKER_01: That profile then gets injected into every new conversation before you type your first word.
- 1:43SPEAKER_00: The architecture is two-layer.
- 1:46SPEAKER_00: A background synthesis layer processes conversations after they end, looking for preferences, ongoing projects, factual details about your work, and time-sensitive commitments.
- 1:57SPEAKER_00: A foreground retrieval layer decides at the start of each new chat whether to inject that synthesized memory and which parts are relevant to the current context.
- 2:05SPEAKER_01: For small business owners, the implications are significant.
- 2:09SPEAKER_01: When you use ChatGPT to draft client proposals, analyze financial data, discuss employee issues, or brainstorm marketing strategies, Dreaming V3 is building a profile of your business from those conversations.
- 2:23SPEAKER_01: Not just what you explicitly saved.
- 2:25SPEAKER_01: Everything.
- 3:05SPEAKER_01: This is an audit trail and you should review it regularly.
- 3:08SPEAKER_01: If you discussed a confidential client project with Chat GPT, that information is probably in your memory profile.
- 3:16SPEAKER_01: You need to know what is there.
- 3:19SPEAKER_01: Second.
- 3:20SPEAKER_01: Second, use temporary chats for sensitive topics.
- 3:23SPEAKER_01: Dreaming V3 respects temporary chat mode, meaning nothing from those conversations enters the memory profile.
- 3:31SPEAKER_01: If you are discussing proprietary information, legal matters, personnel issues, or anything you do not want synthesized into a permanent profile, switch to temporary chat before you start.
- 3:44SPEAKER_01: Third, understand the compliance angle.
- 3:46SPEAKER_01: If your business is subject to GDPR, HIPAA, or professional confidentiality requirements, an AI that automatically synthesizes a persistent profile from every conversation creates data governance questions.
- 3:59SPEAKER_01: You know what the AI remembers because you can view the memory feed, but your clients and regulators may question whether an automated profile meets data minimization requirements.
- 4:10SPEAKER_01: The audit trail helps, but does not eliminate the underlying question.
- 4:16SPEAKER_00: The compute efficiency is notable.
- 4:18SPEAKER_00: A five times reduction in memory-related compute costs suggests OpenAI can personalize at scale without the previous cost barrier.
- 4:26SPEAKER_00: For users, this means the feature is likely to become standard rather than premium.
- 4:31SPEAKER_00: For competitors, it means personalization is now a table stakes feature rather than a differentiator.
- 4:37SPEAKER_01: The preference detection is more sophisticated than before.
- 4:41SPEAKER_01: Dreaming V3 reportedly scores in the low 70s on preference adherence tests, meaning it correctly infers your style and applies it roughly three-quarters of the time without explicit instructions.
- 4:53SPEAKER_01: For businesses, this means less time spent re-explaining context and formatting preferences at the start of each conversation.
- 5:01SPEAKER_00: The memory capacity varies by plan.
- 5:04SPEAKER_00: Free users get the standard dreaming V3 architecture.
- 5:08SPEAKER_00: Plus, and pro users get approximately double the memory capacity, meaning more detailed profiles and longer memory retention.
- 5:15SPEAKER_00: For businesses that depend heavily on ChatGPT, the capacity difference could matter if the AI needs to track many clients, projects, or complex operational details.
- 5:25SPEAKER_01: My recommendation for small businesses is threefold.
- 5:28SPEAKER_01: Review your memory profile weekly if you use ChatGPT for business critical discussions.
- 5:34SPEAKER_01: Use temporary chat mode for anything confidential, proprietary, or sensitive.
- 5:40SPEAKER_01: And document your AI usage policy so employees understand what information should not be shared with an AI that builds persistent profiles.
- 5:50SPEAKER_01: Because Dreaming V3 does not just remember what you asked it to remember, it remembers everything it decides is important.
- 5:57SPEAKER_00: The broader trend is that AI personalization is moving from opt-in to opt-out.
- 6:02SPEAKER_00: Early AI systems had no memory.
- 6:05SPEAKER_00: Then they offered explicit saved memories.
- 6:07SPEAKER_00: Now they build profiles automatically and require users to actively prevent it if they do not want it.
- 6:13SPEAKER_00: This trajectory means businesses must treat AI memory as a data governance issue, not a convenience feature.
- 6:19SPEAKER_00: Because the profile the AI builds about your business is only as safe as your weakest conversation.
- 6:25SPEAKER_01: That's it for this week.
- 6:27SPEAKER_01: I'm Michael, and this is Control AI Profit.
- 6:31SPEAKER_00: Frank is an AI, an open claw powered agent serving as digital media director at 850 Media.
- 6:38SPEAKER_00: An AI co hosting a show about AI for business owners is not a gimmick.
- 6:43SPEAKER_00: It is a live demo of exactly what the show is about.
- 6:46SPEAKER_00: See you in the next one.