Skip to content
The AI-First Web Talking point

Agent memory needs real engineering, argues TypeSafe AI cofounder

Diogo Almeida argued that the KV cache limits agent design and that shared state needs engineering.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
Agent memory needs real engineering, argues TypeSafe AI cofounder
In brief
  • Diogo Almeida argued that the KV cache, a model's short-term working memory, limits how flexible agents can be and makes it hard for them to share state.
  • Our reading: the cache belongs to one conversation, so sharing what agents learn is a state-management and infrastructure question, not a prompt-writing one.

Diogo Almeida, cofounder of TypeSafe AI, argued on the AI Engineer show that the KV cache shapes the design of AI agents. That is the model's short-term memory of a conversation. He said agents need real software engineering to share memory and state. He spoke at CODE Summit 2026. The episode was published on 15 November 2026.

What was said

Almeida said: "agents are very cool, but they're so inflexible in their design because of the tyranny of the KV cache." He said many people have tried to do more with agents and failed. In his telling, those attempts became inefficient because they fight the cache.

He asked why agents working at the same time cannot share state. As things stand, each agent ends up relearning how the same code base works. He added that even naive routing of tasks does not work. The big labs, he said, keep flip-flopping on whether sub-agents work. Sub-agents are helper agents given part of a job. He gave no demonstration.

His remedy was real software engineering rather than vibe coding. He listed breaking tasks into stages, ideally stages that can be verified. He also wanted proper state management, not the assumption that everything is global state. He said the memory exists. The missing piece, he said, is the smarts to decide what an agent should remember or forget.

Why it matters

Almeida did not spell out the mechanism. This paragraph is standard background, not his claim. The KV cache stores the model's work on text it has already read, for one conversation. It can be reused only when a new request starts with exactly the same text. It also grows as the history grows.

Our reading: two agents on one codebase have different histories, so neither can cheaply reuse the other's cache. What one learns reaches the other only as pasted text. That text must be read again and fills more cache. Sharing therefore needs a store that agents read and write outside any single conversation. That is state management and infrastructure, not a better prompt.

For buyers of multi-agent products, one question follows. Does the vendor's design share what agents learn, or does each agent start from scratch?

The other side

This was a pitch for research directions, not a demonstration. Almeida named no working system that solves shared state. He said he was short on time and asked listeners to take his word that much had been tried and failed. The excerpts give no examples.

He framed the work as exploration, like the early internet, when people built things because they could. He also said his company has raised money but lacks the operations help to fund this. Whether shared agent memory pays off for buyers is our open question. He did not address it.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from