Architecture: The Context Engine

The solution is built on three pillars: Memory, Orchestration, and Context.

1. The Context Object

Context is not just text; it is a structured object.

• Layer 1: Identity (RBAC) - Security bounds.
• Layer 2: Episodic State - Short-term conversation history.
• Layer 3: Semantic Knowledge - Long-term facts (Zep/Graphiti).
• Layer 4: Operational State - Workflow status (Temporal).

2. Memory Layer: Zep & Graphiti

We use Zep (Open Source) for the memory layer.

• Temporal Knowledge Graphs: Models how facts change over time (Bi-Temporal Modeling).
• Integration: Zep runs as a service, backed by Postgres.

3. Orchestration Layer: Temporal

We use Temporal to guarantee “Durable Execution”.

• Grid Dynamics Pattern: Decouple the “Brain” (LangGraph) from the “Spine” (Temporal).
• Resilience: If the agent crashes, Temporal retries from the last saved state.
• Human-in-the-Loop: Supports long waits (days/weeks) for approval signals.

4. Ingestion: Unstructured.io

A centralized ETL pipeline processes PDFs, HTML, and Slack dumps into structured data for Zep.