Back to Archive

LA-AI Insights: AI Memory & What’s Really Being Stored

Two very different approaches to remembering your conversations.

Tuesday, October 21, 2025

Share this newsletter:


Sorry for the two-week gap in newsletters. Things got busy, but we’re back.

You’ve probably noticed the headlines. ChatGPT now remembers everything you’ve ever told it. Claude can search through your past conversations. It sounds like AI assistants finally get you. But scratch the surface, and you’ll find something more complicated.

Most people assume your AI is building up this comprehensive understanding of you over time, like a friend who remembers your birthday and knows you take your coffee black. That’s happening, but the how matters more than you think.

When you’re chatting with an AI in a single conversation, it’s using what’s called a context window—basically its working memory for that session. Everything you say stays accessible until you close that chat. Then it’s gone. The next conversation starts from scratch.

The new memory features change this, but ChatGPT and Claude took different approaches. ChatGPT now automatically extracts “memories” from your conversations. It picks out facts it thinks are important and loads those into every future conversation. The system references all your past conversations to deliver more relevant responses. You can view, edit, or delete individual memory entries in your settings, and you can turn off the memory feature entirely if you want. But by default, it’s building that profile.

In April 2025, OpenAI expanded this capability, letting ChatGPT reference all your previous conversations indefinitely. Not just snippets—everything. It’s genuinely impressive from a technical standpoint, and while you maintain control over what’s stored, the system is designed to remember automatically unless you intervene.

Claude went a different direction. When Anthropic rolled out memory in September 2025 for Team and Enterprise users, they built it around transparency and compartmentalization. Claude can recall projects, preferences, and conversations, but users can view, edit, or disable the feature. For business users, there’s an automatic component that can be turned off or restricted by admins. Claude also keeps separate memories for different projects. Your work stuff never touches your personal conversations.

It’s a different philosophy. ChatGPT optimizes for convenience—it wants to feel like it just knows you. Claude optimizes for control and transparency—you decide what it remembers and when, with clear visibility into what’s being recalled.

The data retention story got complicated earlier this year. In May 2025, a federal court ordered OpenAI to preserve all ChatGPT conversations—including deleted ones—as part of The New York Times copyright lawsuit. For months, people who thought they’d permanently erased sensitive conversations discovered those chats were still on OpenAI’s servers. Enterprise customers and API users with special agreements were exempt. Regular users weren’t.

OpenAI called it a “privacy nightmare” and appealed. On October 9, 2025, the judge lifted the broad preservation order.

OpenAI can now delete most user data again, except for logs linked to plaintiff-flagged accounts. The temporary freeze is over, but it revealed just how much data these systems hold and how little control users had during that window.

So what’s AI actually remembering? In Claude’s case, your exact words from past conversations, retrievable on demand with clear transparency about when and how. In ChatGPT’s case, an ever-growing profile of facts, preferences, and patterns extracted from everything you’ve said, with controls you can adjust if you know where to look. Both are powerful. Both are useful. But they handle your data very differently.

The next time an AI “remembers” something about you, you’ll know what’s really happening—and you can make an informed choice about what you’re comfortable having it store.



Upcoming Events

Our October Mobile meetup lands on Halloween this year, so we decided to shake things up. Instead of our usual afternoon gathering, we’re hosting a special Halloween Lunch & Learn at Innovation Portal (lunch included). Same great content, different time, bonus spooky vibes.Costumes are absolutely welcome but not required. Come as you are, or come as a friendly robot. Your call.

Invalid Date

Innovation Portal (358 St. Louis St., Mobile, AL

This Week in AI

1. Claude Code gets a web version—but it’s the new sandboxing that really matters

Anthropic's Claude Code introduces critical sandboxing capabilities that enable secure code execution within AI environments, representing a significant advancement in enterprise AI safety. The sandboxing technology addresses fundamental security concerns that have limited AI code generation deployment in production environments. This development signals a maturation of AI development tools from experimental to enterprise-grade solutions, potentially accelerating adoption across organizations previously constrained by security requirements and establishing new industry standards for safe AI-assisted development.

Ars TechnicaRead more

2. AI researcher Andrej Karpathy says agentic AI is years away from matching industry hype

Former OpenAI and Tesla AI director Andrej Karpathy warns that autonomous AI agents remain years behind current market expectations, challenging widespread industry optimism about near-term agentic capabilities. His assessment suggests significant technical hurdles persist in developing truly autonomous AI systems capable of independent task completion. This reality check from a respected AI researcher signals potential market recalibration and strategic planning adjustments across enterprises betting heavily on agentic AI deployment timelines.

The DecoderRead more

3. Google AI tool pinpoints genetic drivers of cancer

Google develops AI system capable of identifying specific genetic mutations driving cancer development, representing significant advancement in medical AI applications. The tool demonstrates AI's growing capability in complex biological pattern recognition and diagnostic support for oncology research. This breakthrough showcases AI's expanding role in healthcare innovation and precision medicine, potentially accelerating cancer research and therapeutic development. The advancement signals broader trends in AI-powered medical discovery and diagnostic enhancement across healthcare systems.

AI NewsRead more

4. OpenAI builds 'AI for Science' team to advance computational discovery

OpenAI establishes dedicated scientific research division focused on computational discovery and breakthrough research applications. This strategic move positions OpenAI beyond consumer AI into high-stakes scientific computing, potentially competing with Google DeepMind's scientific initiatives. The team aims to accelerate scientific discovery through advanced AI models, representing a significant shift toward specialized, domain-specific AI applications that could transform research methodologies across multiple scientific disciplines and establish new revenue streams in academic and pharmaceutical markets.

The DecoderRead more

5. Hugging Face launches Omni Chat, an AI router for open source models

Hugging Face introduces Omni Chat, an intelligent routing system that optimizes queries across multiple open-source AI models based on task requirements and performance characteristics. This platform democratizes access to specialized AI capabilities by automatically selecting optimal models for specific tasks, reducing complexity for developers and organizations. The launch strengthens Hugging Face's position as infrastructure provider for open-source AI while potentially challenging proprietary model dominance. This development could accelerate open-source AI adoption by simplifying model selection and deployment decisions for enterprises and developers.

The DecoderRead more

6. Anthropic turns to ‘skills’ to make Claude more useful at work

Anthropic launches Skills functionality enabling Claude to automatically select specialized prompts for workplace tasks, marking a significant step toward autonomous AI agents. This development transforms Claude from a conversational AI into a contextually-aware workplace assistant that can adapt its behavior based on task requirements. The launch represents enterprise AI evolution from manual prompt engineering to intelligent task automation, potentially accelerating AI adoption across knowledge work sectors and establishing new standards for AI workplace integration.

The VergeRead more

7. Google Deepmind's new Gemma model helped discover a new cancer treatment pathway.

Google DeepMind's Gemma model contributed to identifying a novel cancer treatment pathway, demonstrating AI's capability in accelerating medical research and drug discovery. The breakthrough showcases how specialized AI models can process complex biological data to uncover therapeutic targets previously invisible to traditional research methods. This application validates AI's potential to transform healthcare research timelines and methodology, establishing precedent for AI-driven scientific discovery in critical medical applications.

The DecoderRead more

Know someone who would enjoy this newsletter?

Forward this email or share the link below

https://la-ai.io/newsletter/view/2025-10-21
Subscribe to LA-AI Newsletter

Join Our AI Community

Get weekly insights on AI innovations and exclusive updates on LA-AI events

Subscribe Now