Cortex (Learning & Memory) Integration

The Cortex is the system-wide memory and learning layer for the cyber-biological architecture. Rather than treating each execution in isolation, the orchestrator leverages Cortex to transform individual agent errors into shared, system-wide capabilities.

Pillar 2: Trajectory Consolidation

A key feature of the Cortex integration established in the system architecture is Trajectory Consolidation. While standard refinement captures error-solution pairs, Trajectory Consolidation memorizes entire successful sequences of tool calls.

When an execution completes successfully, the Orchestrator extracts the sequence of ToolInvocationCompleted events and converts them into a TrajectoryPattern. These patterns are indexed by the task's semantic embedding.

Future agents assigned similar tasks will receive these successful trajectories as "Suggested Workflows" in their system prompts. Over time, frequently successful trajectories are crystallized into permanent Macro-Tools (ToolWorkflows) in the SEAL Gateway, effectively allowing the swarm to learn new procedural skills autonomously.

The Role of Cortex

When an agent encounters an error or requires refinement during the 100monkeys iterative loop, the orchestrator indexes these events. Cortex analyzes the execution history to extract Patterns (error-solution pairs) and organizes them into higher-level Skills.

By connecting to Cortex, your orchestrator can:

• Retrieve known solutions: Automatically provide agents with proven fixes for specific error signatures before they attempt to solve them from scratch.
• Contribute to collective intelligence: Send successful refinement trajectories back to the central graph, improving the system for all connected nodes.
• Enable holographic memory: Access a weighted graph of solutions that dynamically updates based on the continuous success and decay of patterns.

Connecting to Cortex

The orchestrator communicates with Cortex over gRPC and HTTP interfaces defined in the shared API contracts. You configure the connection in your aegis-config.yaml file.

# aegis-config.yaml
spec:
  cortex:
    grpc_url: "env:CORTEX_GRPC_URL"
    api_key: "env:CORTEX_API_KEY"  # Required for Zaru SaaS; absent = memoryless mode

Authentication and Credentials

Access to the central Cortex requires an API key provided by 100monkeys. To maintain the Keymaster Pattern security guarantee, this key should be stored securely in OpenBao (or injected via environment variables) and NEVER placed directly in plain text in the configuration file or exposed to agent runtime environments.

Memoryless Mode

If you do not have a Cortex API key, or if you prefer to run the orchestrator in an entirely isolated, air-gapped environment, AEGIS will operate in Memoryless Mode.

# aegis-config.yaml
spec:
  cortex:
    grpc_url: null  # omit or set null to run in memoryless mode

Behavior in Memoryless Mode

When Cortex is disabled:

• No Pattern Retrieval: Agents will rely entirely on their intrinsic foundational model intelligence and whatever context is provided in their base prompts. They will not receive injected "hints" or known solutions from previous runs.
• No Telemetry Export: The orchestrator will not transmit any execution trajectories, error signatures, or successful code diffs externally.
• Standard Iteration: The 100monkeys loop (Execute → Evaluate → Refine) functions normally, but the loop starts from a blank slate on every execution.

Memoryless mode is fully supported for local development, highly sensitive enterprise deployments, and open-source hobbyists.

Discovery Indexing

When a Cortex connection is configured, the orchestrator automatically forwards all agent and workflow lifecycle events to the Cortex service for indexing. The aegis.agent.search and aegis.workflow.search MCP tools are served through Cortex using semantic vector search.

Without a Cortex connection, discovery tools return a service-unavailable error.