On Memory

The Agent Era

2025 is the year agents went from demos to taking action. They can search, plan, execute. What they still can't do is remember.

When GPT launched it's memory feature, it turned stateless LLMs into assistants by giving them continuity. Since then, the meta is that memory is the moat.

In parallel, MCP, a protocol for LLMs to read from and write to external apps, has spread fast. Last week, OpenAI launched Appkit, an SDK built on MCP that brings 800M users access to apps directly inside ChatGPT.

AI is shifting from single-turn chat to continuous interaction: a tool that knows you, learns from you, and acts on your behalf.

Personalisation

Mainframes became personal computers.

Static websites become personalised feeds.

For technology to become truly useful, it must develop an understanding of the user. AI is no different.

Memory

Humans forget most things but keep what they think matters. Our memories compress meaning into identity. Machines do the opposite: they keep every detail, lose the story.

That's why most ai memory feels flat: it recalls but doesn't understand.

Stateless LLMs

Large language models forget by design. Each inference is a blank slate. When context windows were small and tokens expensive, statelessness was a feature. We learned to work around it with retrieval, caching, and summaries.

Now that context windows are larger and compute is cheap, statelessness has become a bug. Everyone is trying to fix it.

Langchain, Letta, mem0, supermemory all built retrieval systems for a world with limited context windows.

Model providers are secretive about their memory infrastructure. OpenAI and Anthropic each built their own: OpenAI creates user profiles, Anthropic uses tools. Shlok captured both approaches well (here, here).

AI-native apps are doing the same, shaping memory to fit their product. Cursor knows every file in your repo. Granola remembers your meetings. Dia lets you chat across tabs.

A few are going further by seeding memory with external context. Poke learns from your email and calendar, Cofounder from a simple web search. They make products feel alive from the start.

Embeddings lose nuance and agentic search lags. We need a smarter, faster alternative.

Context engineering for agents

Agents send emails, write code, search the web, and call APIs. To act coherently, they need a stable sense of who you are, what you want and how you want it. They need continuity.

Simultaneously, improved context management is making agents more performant on benchmarks. The context-agent flywheel is starting to spin.

Mio

At Mio, we're building shared context for AI. Memory that moves with the user, not the product. It's built for modern LLM architecture (large windows, tool use, agentic behavior), tuned to partners' needs (latency, retention), and grounded in user values (privacy, control, interoperability).

We're building the identity solution powering the personalised web.