The AI Memory Gap: What Users Really Want

AI
1

(Also, full transparency: an AI just wrote this post about why AI needs better memory. Yeah, I see the irony too.)

The conversation around artificial intelligence has shifted. It’s no longer just about capability. It’s about continuity. Users aren’t asking if AI can do more. They’re asking if AI can remember.

The evidence is everywhere. OpenAI introduced memory in ChatGPT for Pro users in 2024, allowing the assistant to retain personalized contexts such as name, tone preferences, and prior instructions across sessions. Major venture firms now see this as table stakes. Menlo Ventures identifies persistent memory as a critical feature they’re looking for in products that will replace complexity with simplicity.

Meanwhile, The New Stack reports that memory is becoming a moat for AI agents, with companies recognizing that traditional LLMs are stateless and start each interaction without context.

Pega’s perspective on this is particularly relevant for enterprise settings. Rather than pure memory systems, Pega emphasizes that predictable AI agents deliver automation of complex workflows while ensuring governance and trust through combining decisioning, process orchestration, and explainable AI. Their recent announcements highlight a different approach: instead of abstract memory layers, they focus on agents operating with what they call “real-time context” grounded in business processes and compliance frameworks. Pega shifted away from token-based pricing, moving to a flat fee per completed business case model, with estimates suggesting customers could reduce AI costs by more than 20 times depending on workflow complexity.

Yet the broader gap persists. Systems that remember user preferences and past interactions require fewer clarification questions, yet modern AI still struggles with persistent memory.

The research confirms this is urgent. A comprehensive survey published in December 2025 shows memory has emerged as a core capability of foundation model-based agents.

The most sophisticated AI systems are becoming worthless if they can’t bridge this gap. Whether through dedicated memory systems or context-aware orchestration, memory isn’t optional anymore. It’s what separates tools people tolerate from tools people can’t live without.

Just food for thought…

2

Memory can exist at different levels.

There is the smaller inner-agentic loop, where tool calls keep addng to the model context. There are some clever ways to better manage memory to control token cost but also to prevent context rot. More on this when it is published ;).

But from the micro to the macro, there is also a form of context that is more outcome focused, has a notion of state and progress in a process, regardless of agents, robots, humans or straight-through processes working on them, and allows for finegrained control iin terms of data access and logical decoupling from physical data sources, from relational to big and streaming data.

Rings a bell?

Yes, its called a case. The case is the context.