Welcome to VersarAI Velocity. Chances are — you didn’t know you had signed up for this newsletter. That’s ok, we love you and wanted to share things with you that we are so excited about! Don’t worry, you can opt out if you don’t feel that love or excitement! Three things this time: a take, a tool, and a tease.

That's right — it's time to move on and recognize that while larger context windows and memories are helpful, they can't meet the needs for modern personalization.

Context is a scratchpad. Memory is a filing cabinet. Neither one teaches an AI how you think.

Stuffing more facts into a prompt doesn't make a system understand you. It just makes it slower and more confident about the wrong things.

What's missing is structure and that’s where we have introduced a new concept in Agentic AI — Sim. A Sim is a compact model of how a person reasons in a given context — their preferences, constraints, and decision patterns, not their transcript.

That's the shift. From recall to reasoning. From history to perspective. Read the paper conceptualizing and explaining this in IEEE Computer Magazine.

Today we're releasing SimBench 1.0 — the first benchmark built specifically for cross-context personalization.

Existing benchmarks miss the point. LaMP retrieves facts. LOCOMO tests memory recall. τ-bench measures tool use. None of them ask the real question: can a model give the right answer when a user's contexts pull in opposite directions?

SimBench does. 47 tasks across 9 users, each one designed so the correct plan depends on which Sims are active and how their constraints arbitrate.

We ran it across GPT-4o, Claude, and Llama. Procedurally grounded Sims beat RAG by ~8 points on GPT-4o. The pattern holds across models.

Grab it, fork it, break it: github.com/versarai/simbench

Every platform needs two halves that need each other. Ours are SimCore and Aria.

SimCore is the MCP server. It builds Sims, routes between them, resolves conflicts, and ships the result wherever an agent needs personalization — SDK, MCP server, or hosted endpoint. It's what other companies plug into.

Aria is the client. A working iOS app that uses SimCore the way a real customer would, on a real device, with real constraints. It's how we prove the platform works before we ask anyone else to bet on it.

One without the other doesn't make sense. A server with no client is a thesis. A client with no server is a demo.

Together they're a product.

Interested in seeing how this could accelerate your Agentic AI advancements or provide easy and effective personalization? Contact us!

More soon. Or sooner. In the meantime, follow along: X/Twitter · LinkedIn