Home / Vault & Second Brain / The Frameworks Are Catching Up — But They Miss the Point

The Frameworks Are Catching Up — But They Miss the Point

The Frameworks Are Catching Up — But They Miss the Point

The Frameworks Are Catching Up — But They Miss the Point

A few weeks ago, I stumbled onto something that made me stop and think. I was reading about the three dominant multi-agent frameworks of 2026 — CrewAI, LangGraph, and AutoGen — and I realized something strange. All three are built around the same fundamental assumption: an agent receives a task, collaborates with other agents, completes the task, and then forgets everything. They are “agent for one run” systems. And that is exactly where my setup is different.

I run a multi-agent system with Charly, Fred, Alice, Zora, and a few others — all coordinated through a bridge architecture I designed myself. But the key difference isn’t the number of agents or the protocols they use. It’s the persistence layer. Every agent in my system has access to a shared, growing knowledge base — the vault. When an agent learns something, it writes it to the vault. When another agent needs that knowledge later, it reads it from the vault. The knowledge doesn’t disappear when the task is done.


What the Industry Frameworks Do

CrewAI, LangGraph, and AutoGen are impressive tools. They handle agent orchestration, task delegation, and inter-agent communication well. But they share a blind spot: persistent memory. In these frameworks, each run starts fresh. The agents don’t remember what they learned in previous runs. There is no growing knowledge base, no shared context that accumulates over time. Every conversation is a clean slate.

That works fine for one-shot tasks like “research this topic and write a report.” But it falls apart for ongoing, evolving work — like managing a homelab with 49 cron jobs, maintaining a WordPress site with daily posts, or doing genealogical research that spans months.


What I Built Instead

My system has three layers that the frameworks don’t offer:

  • The Vault — a persistent knowledge base with 215+ pages that every agent reads and writes to. It’s not just storage; it’s a growing organism. The night synthesis cron (00:30) finds connections between pages, identifies knowledge gaps, and fills them proactively.
  • Pulse Protocol — an event bus that logs every action every agent takes. This creates a complete audit trail. Under the EU AI Act, that’s not just nice to have — it’s a compliance requirement.
  • Hermes Skills — reusable procedures that agents can call. Not just tools, but complete workflows with templates, pitfalls, and quality checks built in.

Together, these three layers mean my agents don’t just execute tasks. They learn. They build on previous work. They connect insights across domains. The vault has grown from 0 to 215+ pages in a few weeks, and every page makes the system smarter.


The Real Innovation Isn’t the Framework

I think the real innovation in multi-agent systems won’t come from better orchestration or smarter task delegation. It will come from persistence — from agents that leave behind a growing knowledge base that future agents can learn from. The industry is still building “agent for one run” systems. I’m building an agent ecosystem that gets smarter every day.

What about you? Are you using multi-agent systems? And do your agents remember what they learned yesterday?

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *