AGINUS: Arvor's next agent, and why it's actually going to remember you

I spent two decades chasing one specific problem: getting systems to find the right piece of information inside billions of documents. Job and résumé search at Catho, enterprise search across banking, retail and telecom after that, data and AI at Semantix. Across all those years, the pattern that bothered me most was never search relevance — it was memory. Systems that answer well once and forget everything the next time you talk to them.

2026’s AI agents inherited that flaw. They’re brilliant for 200,000 tokens and then reset to zero. You re-explain the context, re-import the history, repeat the same correction for the fifth time. That isn’t intelligence. It’s amnesia with good delivery.

That’s the specific problem AGINUS is going after. This post is an honest teaser: the project is in progress, there’s no promised ship date, and I’ll tell you exactly where it stands today.

What AGINUS is

AGINUS is Arvor’s multi-purpose agent — built to talk inside the tools you already use, not inside one more isolated browser tab. Its first surface is Telegram, because that’s where most people already spend part of their day: quick messages, real-life context, no friction from opening a new app.

The goal isn’t another generic chatbot. It’s an agent that accumulates context about you — projects, preferences, past decisions — and uses that to respond better each time, instead of starting from zero every time the context window runs out.

Memento as the backbone

What separates AGINUS from an LLM wrapper with a pretty prompt is the memory architecture underneath: Memento, Arvor’s local-first memory system, already in active development.

Local-first memory means your data lives on your machine or your own infrastructure first, not scattered across a third-party provider by default. And “with provenance” means every fact the agent remembers carries its origin: when it was said, in which conversation, with what confidence. That matters because an agent that silently “remembers wrong” is worse than one that remembers nothing at all — it makes mistakes with conviction. Provenance is what makes memory auditable and correctable, not just something that accumulates.

That backbone is already being stress-tested on another front internally: aginus-v2, an agent gateway built on top of the Memento backbone. That’s where we validate the architecture holds up under real use before we expose it externally.

Why this is 2026’s frontier, not just a feature

Persistent memory across sessions has stopped being a research niche and turned into a category fight. Hermes Agent, from Nous Research, is the most visible example: an agent that accumulates memory across sessions and writes its own skills as it learns, with real community traction since launch. It’s not the only one, but it’s telling — the market noticed the agent bottleneck isn’t reasoning, it’s continuity.

Arvor isn’t trying to replicate that movement. We’re solving the same problem through the lens of someone who built retrieval infrastructure in production — Solr, Lucene, Hadoop, Spark — before “agent” was a product word at all. Data provenance and reliability aren’t new to us. They’re the craft.

Where this lands, and the invitation

I won’t promise a date. I promised too many dates earlier in my career, and I learned not to do that in public. What I can tell you: AGINUS is being built in public, and the next updates will land here, straight, with what worked and what didn’t.

If you want early access, or you already feel the pain of agents that forget everything between one conversation and the next, get in touch and join the waitlist. And if you need this problem solved right now, with what we already have running in production for clients, talk to us about agentic consulting — not the same product, but the same thinking behind it.

The tree remembers where it came from. That’s how it knows where to grow next.