Skip to main content

​
Introduction to Egregore

Putting the “Engineering” in Context Engineering Egregore is an AI agent framework that revolutionizes how developers work with LLM context windows. Built on the PACT (Positioned Adaptive Context Tree) architecture, Egregore treats context like a DOM, enabling precise manipulation and control.

​
The Problem

When OpenAI released ChatGPT and its API, they introduced the current primitives for context manipulation: the simple user/system/assistant chat log system. As developers built more sophisticated agents, they discovered this append-only structure has critical limitations:

​
Real-World Example

Consider a typical CLI tool workflow:
  1. LLM reads file A (100 lines)
  2. LLM edits file A (now 112 lines)
  3. LLM does something else
  4. LLM tries to edit file A again → fails (doesn’t know current structure)
  5. LLM re-reads the entire file
  6. LLM successfully updates the file
This repetitive execution wastes tokens, pollutes the context window, and makes unnecessary API calls. There had to be a better way.

​
The Solution: PACT Architecture

Egregore treats the context window like a virtual DOM. Instead of an append-only chat log, you get:
  • Precise positioning - Address any component with coordinates like (0,1,0,-2,1)
  • Dynamic manipulation - Insert, update, delete components anywhere
  • Temporal management - Components can expire after N turns or cycle periodically
  • Stateful views - LLMs maintain awareness of current state without re-reading
Think of it like HTML/CSS for your context window. You can query, position, and manipulate components with the same precision you’d use for DOM elements.

​
Key Features

​
🎯 PACT v0.1 Compliance

Full specification compliance with automatic serialization. Context components inherit from PACTNode with canonical fields for universal addressability.

​
🔄 Component Lifecycle System

Four component types with TTL (Time To Live) and cadence:
  • Permanent - Never expire (default)
  • Temporary - Expire after N turns
  • Sticky - Expire and reappear at same position
  • Cyclic - Expire and reappear on intervals

​
🏗️ Context Scaffolds

Persistent agent memory and capabilities. Scaffolds auto-generate tools from methods using the @operation decorator and communicate through a formal IPC system.

​
🪝 Comprehensive Hook System

Decorator-based lifecycle hooks for:
  • Tool execution (pre_call, post_call, on_error)
  • Context operations (before_change, after_change)
  • Streaming (on_chunk)
  • Messages (on_user_msg, on_provider_msg)
  • Scaffolds (on_state_change)

​
🔌 Universal Provider Interface

30+ AI providers with standardized API:
  • OpenAI, Anthropic, Google, Mistral, Groq, and more
  • Universal token counting
  • OAuth interceptors for premium models
  • 3-tier parameter processing

​
🌊 Workflow System

Node-based execution with:
  • Native agent discovery
  • Parallel processing
  • Decision nodes
  • Type checking and validation
  • Dictionary-based result aggregation

​
📊 Telemetry & Analytics

Built-in observability:
  • Context snapshots
  • Component change tracking
  • Tool execution monitoring
  • JSONL export for analysis

​
The Egregore Advantage

  1. More Steerable LLMs - Precise control over what the model sees and when
  2. Token Efficiency - No redundant re-reading of content
  3. Context Hygiene - Expire unnecessary components automatically
  4. Stateful Views - Maintain “viewports” for dynamic data
  5. Reactive Updates - Scaffolds automatically re-render on context changes

​
Architecture Overview

Agent
  ├── Context (PACT Tree)
  │   ├── SystemHeader (depth -1)
  │   ├── ActiveMessage (depth 0)
  │   └── History (depth 1+)
  │
  ├── Scaffolds (Persistent Memory)
  │   ├── InternalNotes
  │   ├── FileManager
  │   └── Custom Scaffolds
  │
  ├── Tools (Auto-generated)
  ├── Hooks (Lifecycle Observers)
  └── Provider (30+ AI Models)

​
Next Steps

Quickstart

Get up and running in 5 minutes

Core Concepts

Learn the fundamentals of PACT

Examples

See working code samples

API Reference

Complete API documentation

​
Version

Current version: v0.3.5a3 Last updated: October 2024