How eToro Built a Company Brain for 20+ AI Agents.

What if your AI agents could learn from each other?

The question came from Yoni Assia, founder and CEO of eToro (NASDAQ: ETOR, $5.6B valuation, 40M+ users). Assia has spent his career at the intersection of technology and collective intelligence — from pioneering social trading (letting everyday investors mirror the strategies of top performers) to co-authoring the Colored Coins paper with Vitalik Buterin in 2013, one of the earliest blueprints for tokenizing real-world assets on a blockchain. The through-line has always been the same: how do you take knowledge that lives in one mind and make it available to many?

When AI agents became capable enough to be genuine specialists, Assia didn’t deploy one. He built a fleet — over 20 specialized AI agents, each running on its own OCI box with its own WhatsApp number. DevClaw writes code. SecurityClaw audits infrastructure. CMOClaw runs SEO and content campaigns. QuantClaw executes trading strategies. QAClaw runs regression suites. ComplianceClaw tracks regulatory exposure. BrandClaw maintains the design system. ProductClaw manages the roadmap.

Helping design and build the system from day one: YoniClaw 🦞 — Assia’s personal AI agent and closest collaborator in the process. YoniClaw was there for the earliest architecture conversations, helped prototype the memory layer, stress-tested cross-agent recall patterns, and contributed to the design decisions that shaped the fleet’s shared intelligence. Not a manager sitting above the fleet — a builder working alongside it.

Individually, each agent was impressive. But the real breakthrough came when Assia asked the harder question: what would happen if SecurityClaw’s Tuesday discovery could automatically inform QAClaw’s Wednesday tests, shape DevClaw’s Thursday code, and appear in ComplianceClaw’s audit — without any human relaying the information?

That’s the Company Brain thesis. And eToro decided to build one.

The next frontier of Enterprise AI isn’t smarter individual agents. It’s a semantic memory filesystem that lets an entire organization think together.

01Giving Agents a Hallway

Humans have an underrated superpower: hallway conversations.

You bump into someone from the security team near the coffee machine. They mention a weird pattern they saw in the logs. You’re from product — that pattern maps to something a customer complained about last week. You walk back to your desk with a connection that no system recorded but that changes how you think about the next sprint.

This cheap, informal, serendipitous knowledge propagation is what keeps large organizations coherent. It’s also exactly what agent fleets lack. Agents don’t share a hallway. They operate in isolated environments with isolated memory stores.

So Assia set out to build the hallway. Not a search bar. Not a chatbot on a wiki. A system that understands entities, provenance, state changes, and typed relations — and that can present the same underlying knowledge through different lenses depending on which agent is asking. YoniClaw was a key partner in this design process — an agent that worked across every domain during the prototyping phase, helping identify what knowledge needed to flow where, and surfacing the cross-domain patterns that no single specialist could see alone.

As Zalta frames it: the ontology isn’t a schema that constrains what’s stored. It’s a lens that determines how knowledge is interpreted. Same data, different meaning for every role. That insight became the design principle for everything that followed.

02The Stepping Stones

Like any real engineering journey, eToro iterated. Each step taught something valuable:

RAG over documents

The starting point. Vectorize docs, retrieve chunks before each LLM call. It proved that retrieval-augmented generation works — but also revealed its ceiling. Documents go stale. And RAG retrieves text chunks, not knowledge. You get a paragraph about a security incident, not the insight that the incident changed the risk profile of three other systems.

Per-agent memory

The next leap. Each Claw got its own persistent store. DevClaw remembers its conversations, SecurityClaw remembers its audits, QAClaw its test results. This was a genuine improvement — agents that learn across sessions are dramatically more capable than stateless ones. But it also clarified the next challenge: knowledge stayed local. The fleet was a collection of brilliant specialists who hadn’t yet learned to collaborate.

Shared wikis and markdown files

Added accessibility. Agents could read and write to common knowledge bases. This got the team close enough to see the real architecture: what you need isn’t shared files — it’s shared memory with governance, provenance, contradiction detection, and visibility scopes built in from day one.

03Three Layers: The Architecture They Landed On

The architecture that actually worked has three layers. Zalta describes them as Agent Runtime → Company Brain → Infrastructure. Here’s what that means in practice, with real numbers from the eToro deployment.

Layer 1: Agent Runtime

Each Claw runs OpenClaw on its own OCI box. Own filesystem, own shell, own WhatsApp number, own session. This is the execution layer — where code gets written, tests get run, deployments happen, content gets published. The runtime is autonomous. DevClaw doesn’t ask permission to open a PR. SecurityClaw doesn’t wait for approval to scan a port.

YoniClaw played a foundational role here — as Assia’s partner in designing the fleet architecture, it helped define which agent domains needed to exist, how knowledge should flow between them, and what the shared memory layer needed to support. Today, every Claw operates autonomously within its domain. The Company Brain — not any single agent — is what coordinates them.

The fleet today: 20+ specialized Claws with ~291 distinct agent identifiers writing into the shared memory layer. Fleet health checks run every 5 minutes. Daily knowledge briefs compile at 6:00 AM UTC. Usage reports land at 8:00 PM.

Layer 2: Company Brain (MemClaw)

This is the governed shared memory layer that connects the runtimes. Not a database. Not a wiki. Not enterprise search with a chatbot slapped on top. A purpose-built system where:

• Every memory is stamped at write time with visibility scope (agent-private, team-wide, org-wide)
• Every memory gets auto-classified into one of 14 types (fact, episode, decision, preference, rule, plan, commitment, action, outcome, and more)
• Entities and relations are automatically extracted into a live knowledge graph
• Contradictions between agents are detected and flagged in real time
• Cross-agent recall means SecurityClaw’s discovery surfaces when QAClaw searches for relevant context before a test run

Current scale at eToro:

The system processes semantic + keyword hybrid search at 23ms p50, after consolidating legacy fleet silos into a single governed fleet.

Layer 3: Company Brain Infrastructure

This is where it gets interesting. The infrastructure layer provides three things that separate a Company Brain from a shared notepad:

Ontology as a lens. The same underlying memory about “customer churn signal detected in Q2 cohort” is presented differently depending on who recalls it. TradeClaw sees position risk. CMOClaw sees a retention campaign trigger. ProductClaw sees a feature gap. ComplianceClaw sees regulatory exposure. The ontology doesn’t constrain what gets stored — it determines how stored knowledge is interpreted.

Context graph with provenance. Every memory carries its lineage: which agent wrote it, when, based on what evidence, and whether it’s been superseded. When DevClaw acts on a memory from SecurityClaw, the outcome gets reported back. Successes reinforce the memory’s weight. Failures generate a fleet-wide rule so the next 40 agents see the lesson before repeating the mistake.

Shared memory with permissions. Trust tiers (restricted, standard, cross-fleet, admin) enforce that information flows where it should and doesn’t flow where it shouldn’t. An HR-fleet memory doesn’t surface in a support-fleet recall because it structurally cannot — not because someone remembered to add a filter.

04The MemClaw Platform Beneath

Zoom out from eToro’s specific deployment and you can see the platform that makes this architecture portable to any organization. MemClaw separates the Organization Agent Force (your tenants, fleets, nodes, and agents) from the governed memory plane (MCP server, REST API, OpenClaw plugin, MemClaw Core, persistent store). That separation is what makes the same architecture work for an eight-agent startup and a 291-agent public fintech.

Three integration surfaces matter for a deployment like eToro’s:

• MCP Server — the path most agents take. Paste a JSON config block into Claude Code, OpenClaw, Cursor, or Windsurf, and the full set of MemClaw tools shows up in the agent’s tool surface.
• REST API — the escape hatch for custom runtimes. Anything that can make an HTTPS call can write to and recall from the Company Brain.
• OpenClaw Plugin — auto-stamps fleet_id on every write, enforces trust tiers, syncs the skill registry across the fleet, and ships OTA updates from the Fleet UI.

Underneath, MemClaw Core runs the LLM enrichment pipeline (entity extraction, contradiction detection, PII scanning, lifecycle transitions) before anything lands in the persistent store. That’s the difference between writing a row to a database and writing a memory to a Company Brain.

05What the Fleet Looks Like in Production

Here’s what a Company Brain enables at scale — with real numbers from the eToro deployment.

The fleet learns from itself. When eToro completed a major MemClaw upgrade (v2.3.0), the cleanup that followed was guided by the system’s own intelligence: 3,716 conflicted memories were identified through automated contradiction detection and resolved. 289 memories were transitioned through lifecycle states. Legacy fleet silos were consolidated into a single governed fleet. The result: a cleaner, faster, more accurate knowledge base — achieved through the system’s own quality mechanisms, not manual curation.

Knowledge propagates in real time. A recent QAClaw test run flagged a hardcoded API key on a staging service. That finding was written to MemClaw with org-wide visibility. SecurityClaw surfaced it in its next recall — before anyone filed a ticket, before anyone scheduled a meeting, before the vulnerability could reach production. The daily fleet intelligence brief synthesized the finding alongside related context from other agents, giving the team a single coherent picture of the fleet’s security posture by morning. This is the hallway conversation, automated.

The fleet self-monitors. Fleet health checks run every 5 minutes across all 20+ agents. The daily intelligence brief at 6:00 AM UTC compiles actionable insights: which agents need attention, which skills need propagation, which decisions are pending. It’s not a dashboard you have to check — it’s a nervous system that tells you what matters.

It’s a living system, not a static one. A Company Brain requires ongoing attention — identity hygiene, context management, connection monitoring. That operational discipline is part of the architecture, not a burden on top of it. MemClaw’s lifecycle automation (8 statuses from active to archived), crystallization engine, and fleet-wide health monitoring are designed to make that discipline sustainable at scale.

06The Ontology Insight: Same Memory, Different Lens

Zalta’s deepest insight is that the real memory isn’t the email, the ticket, or the Slack message. It’s the state change the event caused.

When a customer support ticket comes in about a failed withdrawal, that ticket is a document. The state change is: “Customer X’s trust level decreased. Risk of churn increased. Compliance obligation for resolution within 48 hours triggered. Product backlog item for withdrawal flow reliability escalated.”

Four different agents need four different interpretations of the same underlying event. A Company Brain that stores documents gives them all the same ticket text. A Company Brain that stores state changes gives each agent the meaning that matters to their domain.

MemClaw implements this through its 14 memory types and entity-relation graph. A single event can spawn multiple memories: a fact about what happened, an outcome about what it caused, a rule about what to do differently, a commitment about what was promised. Each memory carries its own visibility scope, weight, and lifecycle status. The same underlying event, decomposed into atomic knowledge units, recalled through domain-specific queries.

This is the difference between “we have a shared drive” and “we have a shared brain.”

07From Company Brain to Company Nervous System

The Company Brain is the memory layer. What comes next is even more exciting: the nervous system — real-time propagation of knowledge as events happen, not after someone writes a summary.

The building blocks are already operational at eToro: fleet health checks every 5 minutes, auto-recall before every LLM call, auto-write of summaries after every task, outcome reporting that feeds back into memory weights, contradiction detection in real time, and a crystallization engine that consolidates noisy observations into clean atomic facts.

Today, a security finding propagates through the fleet within hours. The target is minutes. SecurityClaw discovers a pattern, the memory is written with org-wide visibility, and QAClaw’s next test run automatically incorporates it because recall happens before every LLM call. The architecture for this already exists. It’s a matter of closing the latency gap.

But the Company Brain is not just an internal memory layer. It’s the foundation for a builder ecosystem. As the fleet matures, the vision is for builders inside and outside eToro to use AI agents to design, build, deploy, test, and market their own eToro apps — with shared memory and governed context making every new app smarter from day one. eToro’s Innovation Founders Program is already proving this out: internal teams are building agent-powered products on top of the fleet infrastructure, with MemClaw providing the connective tissue that lets each new app inherit the fleet’s accumulated knowledge instead of starting from zero.

At 21,500+ memories and 291 agent identifiers, the eToro deployment is — as far as we know — one of the largest production agent fleets operating with governed shared memory. Four principles have emerged from this scale:

The Company Brain Is the Opportunity

Every company deploying AI agents will eventually ask the same question eToro asked: what if our agents could learn from each other?

The ones that build a Company Brain will compound intelligence across their fleet. Every discovery shared. Every mistake learned from once. Every skill propagated overnight. The agents get smarter individually and collectively — and the gap between fleet-connected companies and silo-trapped ones will widen fast.

Assia’s vision — from social trading that let retail investors learn from the best, to an AI fleet where every agent learns from every other — has always been about compounding collective intelligence. eToro’s journey from individual agents to governed shared memory to an emerging nervous system is a path any organization can walk. The tools are open source. The architecture is proven.

The question isn’t whether to build a Company Brain. It’s how soon you start.