news · 28 September 2026

AI agents gain better memory, Meta launches consumer AI, and musical hallucinations

This week's AI round-up looks at how AI models are getting smarter about remembering-and forgetting-information, an experimental music device, and Meta's latest consumer AI push. We also consider the bigger picture of how much control we should give AI in our products.

In this round-up

AI agents are getting a memory upgrade
AI agents are learning to forget, and why that matters
Meta pushes consumer AI with 'Muse' agent
Engram sampler uses AI to create experimental music
Deciding how much power to give AI in your products
Anthropic CEO to dine with President Trump

AI agents are getting a memory upgrade

Many AI tools, or 'agents', currently forget previous interactions once a session ends. This means you have to re-explain yourself every time, which makes AI feel more like a disposable tool than a helpful assistant. New approaches are being developed to give these agents a memory layer, allowing them to store and retrieve context across interactions. This includes short-term memory for recent chats and long-term memory for user preferences and tasks. The aim is for AI agents to remember past conversations, learn your preferences, and adapt as your goals change over time.

For events professionals, this could mean AI assistants that truly learn your specific event requirements and preferences. Imagine an AI helping to plan your next conference. Instead of starting from scratch every time, it would remember your previous budgets, preferred suppliers, common delegate queries, and even your personal communication style. This would make drafting delegate comms, managing RSVPs, or even pre-event briefings much more efficient, as the AI would build a consistent understanding of your needs.

AI agents are learning to forget, and why that matters

While giving AI agents memory is crucial, an equally important development is teaching them when to forget or update information. Current memory systems for AI can remember facts too well and too long, without knowing if the information is still true. For example, an AI might confidently quote an old refund policy that has since changed, leading to incorrect information. The challenge is to design systems that can assess whether remembered information is out-of-date, superseded, or no longer relevant.

For event organisers, this is vital for accuracy. Imagine an AI managing supplier contracts or delegate information. If it relies on outdated details, such as an old catering menu or an expired promotional code, it could create significant issues. AI systems that can intelligently update or discard old information would ensure that automated tasks, like sending out event updates or processing bookings, are always based on the most current and accurate data, preventing mistakes that could impact your event.

Meta pushes consumer AI with 'Muse' agent

Meta, the company behind Facebook, has unveiled a new consumer AI agent called Muse. Unlike many other major AI companies focusing on enterprise tools, Meta is concentrating on bringing AI features to its consumer products. Muse is described as a 'party-trick type thing' by some, but Meta sees it as a way to integrate AI into people's daily lives. A key question around Muse is whether users will trust Meta with sensitive information, given its business model relies on advertising.

For the events industry, this is less about a direct tool to use, and more about broader AI adoption. As consumer-focused AI becomes more prevalent, event attendees will become more accustomed to interacting with AI in their daily lives. This could increase their comfort level with AI-powered features at events, such as smart chatbots for information, personalised recommendations, or interactive experiences. While Muse itself may not be a game-changer for events, the wider trend of consumer AI could shape attendee expectations.

Engram sampler uses AI to create experimental music

A music startup called Thoughtful Things has launched Engram, a new sampler and groovebox that uses AI to manipulate audio and generate new sounds. It is designed for experimental sound creation, not for producing chart-topping songs. The device runs a 'tiny AI' locally, without internet connection, and the company states its audio models were trained on commercially licensed, open datasets. The founders describe Engram as a 'field recorder for latent space', encouraging users to explore AI 'hallucinations' and unconventional sound design.

For events, this could open up interesting possibilities in sound production and artistic installations. Event production teams might use such a device to create unique soundscapes for immersive experiences, or live performers could integrate its experimental capabilities into their sets. It is a niche tool, but it points to the creative applications of AI beyond productivity, potentially offering new ways to engage audiences through sound at future events.

Deciding how much power to give AI in your products

A key challenge for product managers integrating AI is determining how much autonomy the AI should have. It is not about whether AI 'can' do something, as models are often capable of many tasks. Instead, the focus should be on the cost when the AI makes a mistake, and who bears that cost. A human error is often one-off and explainable, but a miscalibrated AI can make the same mistake thousands of times in the same direction. Key questions to ask are whether a user can see and correct AI mistakes, and if not, how severe the impact of an uncorrected error would be.

For event tech providers and in-house teams developing AI features, this is a crucial consideration. When designing AI for tasks like ticketing, delegate registration, or supplier management, understanding the 'error shape' is vital. An AI drafting delegate emails is one thing; an AI making autonomous financial decisions or critical operational choices is another. Always consider the potential impact of an AI error on your event's reputation, budget, and attendee experience before deploying features with high autonomy.

Anthropic CEO to dine with President Trump

Dario Amodei, CEO of AI firm Anthropic, is set to have dinner with former President Donald Trump. This meeting comes amid a debate about AI safety, with Amodei advocating for caution in AI development, while Trump has expressed different views on the technology. Anthropic has previously had a complicated relationship with Trump's administration, even facing a 'supply-chain risk' designation from the Pentagon due to its attempts to implement guardrails around its technology's use.

This story is less about direct application for event professionals and more about the ongoing political and ethical discussions surrounding AI at the highest levels. It highlights that AI development is not just a technical challenge, but also a matter of policy, safety, and international relations. Staying aware of these broader conversations is important context for anyone using or considering AI, as regulatory changes or public sentiment could influence the availability and responsible use of AI tools in the future.

Sources

By Ed Richards

All news