Cortex
Five phases of thought.
Cortex is the reasoning engine at the heart of Pulse. Every message you send triggers a five-phase cognitive cycle — Observe, Think, Act, Verify, Learn — that transforms a simple request into autonomous, verified action.
Unlike conventional chatbots that generate a single response and stop, Cortex operates as a continuous reasoning loop protected by nine guards (empty response retry, refusal detection, announcement detection, and more). It can chain multiple tool calls, verify its own output, recover from errors, and extract lasting knowledge from every interaction. The loop runs up to 18 iterations — compressing context and extending when needed. For complex tasks, Cortex can spawn up to 10 parallel sub-agents, each with their own OTAVL cycle.
This architecture means Pulse doesn't just answer questions — it solves problems. A single request like "schedule a meeting with John next week" triggers observation of your calendar, reasoning about available slots, action via the calendar API, verification of the created event, and learning about your scheduling preferences for next time.
The Five Phases
Observe
Parse the user's intent, load relevant context from Engram memory, gather integration state, and build a complete picture of what's being asked.
Think
Reason about the best approach. Select tools, plan multi-step sequences, consider constraints, and prepare the execution strategy.
Act
Execute the plan using Arsenal tools: API calls, web searches, browser automation, file operations. Each action produces a result that feeds back into the loop.
Verify
Check the result against the original intent. Did the calendar event get created? Does the email draft match the request? If not, loop back and correct.
Learn
Extract preferences, patterns, and corrections from the interaction. Store them in Engram for future use. Every conversation makes Pulse smarter.
Recursive Feedback
The Learn phase feeds back into Observe. Detected preferences refine future observations. Errors become anti-patterns that prevent recurrence.
Full Autonomy
Cortex can run up to 18 iterations per turn without human intervention. When context grows large, it compresses and extends automatically.
Multi-Agent Orchestration
Spawn up to 10 parallel sub-agents for complex tasks. Each agent runs its own OTAVL cycle with tool policies, priority queues, and automatic result aggregation.
How Cortex processes a request
Message received
Your message arrives via any channel — web, WhatsApp, Telegram, Discord, or Slack. Cortex activates.
Context loaded
Engram injects relevant memories, Vault provides document knowledge, and Oracle adds proactive context like upcoming meetings.
Tools executed
Arsenal provides the tools. Nexus selects the AI model. Cortex orchestrates the execution, chaining multiple tools if needed.
Result verified
The output is checked for correctness. If verification fails, Cortex loops back to the Think phase and tries a different approach.
Knowledge extracted
Preferences, patterns, and corrections are saved to Engram. Anti-patterns from failures are recorded to prevent recurrence.