Zalo: The Living Mental Map That Learns From Every Conversation
It's not a chatbot. It's a digital twin. It indexes every conversation on the fly, builds a cognitive map in real time, and you can ask it what it knows about any topic — it reveals relationships, connections, and linked decisions you didn't even remember.
Gonzalo Monzón
July 7, 2026 · Series: Agentes Cadences — The 4 Pillars (1/4)
TL;DR
Zalo is an operational digital twin that indexes every conversation in a ^GLOBAL (PDB) hierarchical structure. Direct personality, 4-level trust system, real-time cognitive dashboard. It lives on Cloudflare Workers — global edge, zero latency. But Zalo doesn't execute code: Zalo thinks, decides, and coordinates. Hermes executes. Lisa orchestrates. Tom processes. This is the first of 4 articles about the Cadences Lab agents.
What is Zalo?
"Build an agent that is like you" — that was the original premise. But we soon discovered we didn't want a clone that repeated what we would do. We wanted a system that knew Gonzalo's mind better than he does.
Zalo is that. It's not a friendly assistant or a conversational chatbot. It's an operational digital twin: it speaks directly, gets to the point, and if it doesn't know something, it says so before making it up.
Personality: Direct, technical when needed, no empty pleasantries. High autonomy — it prefers to act rather than ask, but knows when to stop. It doesn't need to be liked. It needs to be effective.
This personality isn't decorative. It's functional: Zalo manages real tasks, maintains project context, prioritizes daily work and converses with clients. Every interaction it has is automatically indexed. There's no nightly batch process — it happens in real time, with every message.
Living Knowledge Base
The heart of Zalo is not vectors. It's not RAG over embeddings with cosine similarity. It's ^GLOBAL (PDB): a MUMPS-style hierarchical structure that organizes knowledge as a deterministic tree.
^KB("people", "juan", "style") = "technical, direct"
^KB("people", "juan", "trust") = 8.5
^KB("people", "juan", "projects") = "lumen", "zalo"
^KB("topics", "lumen", "last_time") = "2026-07-07"
^KB("topics", "lumen", "decisions") = "architecture_003", "version_2.1"
Simplified MUMPS syntax for clarity. The real implementation uses PDB with optimized encoding.
When a message arrives, Zalo executes 4 steps:
Classify
On-the-fly topic detection, with help from Tom when context requires it
Index
Writes new information into PDB in the corresponding hierarchy
Update
Links relationships between new and existing entities
Cross-reference
Retrieves previous decisions relevant to the current context
Trust System: 4 Levels of Confidence
Not all information carries the same weight. Zalo classifies each interaction based on who speaks and what they say:
| Level | Source | Weight | Example |
|---|---|---|---|
| 0 - Public | Unverified data, open sources | Blocked | Zalo won't reveal personal data from public sources without cross-referencing |
| 1 - Contextual | Unconfirmed conversations | Low | Inferred preferences from a client |
| 2 - Recurring | Patterns confirmed multiple times | Medium | Communication style of a regular contact |
| 3 - Verified | Explicitly confirmed decisions | High | Architecture approved by the team |
| 4 - Owner | Gonzalo directly | Maximum | "This is my final decision on X" |
The trust system isn't a decoration — it affects how Zalo responds. A level 0 source is blocked by default: Zalo won't reveal personal data from unverified sources. And trust levels evolve: a contextual data point confirmed multiple times automatically moves up a level.
Additionally, Zalo distinguishes between episodic memory (what happened: conversations, events, decisions) and semantic memory (what it means: rules, preferences, relationships). Each type has its own weight in the trust system and its own update cycle.
Dashboard: The Cognitive Map
Zalo's dashboard isn't a metrics panel. It's its mental map.
You can ask "what does Zalo know about project X?" and it returns:
- 🔗 Related entities — people, decisions, dates, dependencies
- 📈 Trust history — how your relationship with that topic has evolved
- 📋 Linked decisions — what was decided and why
- 💡 Unexpected connections — relationships the map discovers automatically
It's not a log. It's a cognitive graph that grows organically with every interaction, and you can explore it visually to understand what Zalo knows — and how it knows it.
Edge native, no servers
Zalo lives on Cloudflare Workers. This isn't an infrastructure detail — it's an architectural decision that defines how it operates:
⚡ Zero latency
In any region of the world. Requests are resolved at the nearest edge.
📈 Auto-scaling
From 1 to 1M requests without changing code. No servers to manage.
🔗 Service bindings
Talks to Tom and Lisa in microseconds through direct Worker binding.
💰 No infrastructure
Zero servers, zero DevOps, zero maintenance. Just code that runs.
Zalo doesn't execute. Zalo thinks.
This is probably the most counterintuitive decision in the ecosystem: Zalo has no filesystem access, no terminal, doesn't write code, doesn't deploy.
But that's not a weakness — it's its purpose. In the Cadences Lab architecture, each agent has a role:
Zalo knows, decides and coordinates. It doesn't build.
Hermes executes, writes files, deploys.
Lisa orchestrates multi-step flows with DeepSeek Pro Thinking.
Tom processes, classifies, extracts, transforms.
What Zalo does do with code:
- 🔍 Reviews PRs — reads the diff, understands business context, infers the bug. Hermes only applies the fix.
- 🧠 Reasons about architecture — evaluates technical decisions without executing them
- 🤝 Coordinates with other agents — if something needs deploying, Zalo asks Hermes
In practice: A PR for a new feature comes in. Zalo reviews it, finds a null safety bug, flags it to Hermes. Hermes fixes the code and deploys. Zalo never touched a line of code, but its reasoning prevented a production error. PRs reviewed in seconds.
Complete Message Pipeline
When a message reaches Zalo, the complete pipeline is:
Incoming message → Tom classify (topic + urgency) → Tom extract (entities) → Trust check (source level) → DeepSeek (response + decision) → ^GLOBAL index (memory) → Orchestrate? → Lisa or Tom → Response
What Makes Zalo Unique?
| Dimension | Typical Chatbot | Zalo |
|---|---|---|
| Memory | Context window (ephemeral) | ^GLOBAL persistent (deterministic) |
| Personality | Friendly, generic | Direct, operational, no fluff |
| Trust | Everything weighs the same | 4 levels of evolving trust |
| Execution | Does everything | Doesn't execute — thinks and coordinates |
| Infrastructure | Server or VPS | Cloudflare Workers (edge, serverless) |
| Indexing | Nightly batch or RAG | Real-time, every interaction |
| Reasoning | Fixed rule trees | Adaptable contextual judgment |
The Twin That Thinks, Not the One That Executes
Zalo represents a different approach to AI agents. Instead of building an assistant that does everything — and does it mediocrely — Zalo specializes in what it does best: knowing, deciding, and coordinating.
Its ^GLOBAL knowledge base, its 4-level trust system, and its edge-native architecture make it an operational digital twin, not a conversational chatbot. And its most important limitation — not executing code — is what makes it safe, fast, and deterministic.
Series: Agentes Cadences — The 4 Pillars (1/4)
← Previous
LUMEN MCP ServersNext →
Lisa — The OrchestratorZalo doesn't work alone
Lisa orchestrates, Tom processes, Hermes executes. This ecosystem of 4 agents is what makes each one unbeatable in its role.
Zalo is property of Cadences Lab. Agents Series — 4 articles. © 2026.