Scaffolds Overview

Scaffolds are Egregore’s system for persistent agent memory and capabilities. They’re dynamic components that automatically render into context, provide tool operations, and maintain state across interactions.

What are Scaffolds?

Think of scaffolds as persistent plugins for your agent that:

Maintain state - Remember information across conversations
Provide tools - Expose operations the agent can use
React to changes - Automatically update when context changes
Communicate - Share state through formal IPC system

Scaffolds are like “smart memory” - they observe context, maintain persistent state, and provide capabilities the agent can access.

The Scaffold Pattern

Traditional approaches require manual state management:

# Manual state management - fragile
notes = []

def add_note(note: str):
    global notes
    notes.append(note)
    # Manually insert into context
    agent.context.pact_insert("d0, 1, 0", TextContent(f"Notes: {notes}"))

Scaffolds solve this with automatic state management:

# Scaffold - automatic and persistent
agent = Agent(provider="openai:gpt-4", enable_scaffolds=True)

agent.call("Remember: user prefers Python")
# InternalNotesScaffold automatically:
# 1. Captures the note
# 2. Stores it in persistent state
# 3. Renders it into context
# 4. Provides tools for retrieval

# Later...
agent.call("What do I prefer?")
# Agent has access to all notes automatically

Built-in Scaffolds

Egregore includes three powerful built-in scaffolds:

1. InternalNotesScaffold

Automatic note-taking and memory:

agent = Agent(provider="openai:gpt-4", enable_scaffolds=True)

# Agent automatically captures important information
agent.call("My API key is sk-abc123")
agent.call("I prefer dark mode")
agent.call("My email is alice@example.com")

# Access notes
notes = agent.scaffolds["notes"]
print(notes.state.notes)
# ["API key: sk-abc123", "Preference: dark mode", "Email: alice@example.com"]

# Agent can recall notes
agent.call("What's my email?")
# "Your email is alice@example.com"

Features:

Automatic note capture
Persistent storage across sessions
Search and retrieval operations
Category organization

2. FileManager

Track file operations and maintain file system context:

agent = Agent(provider="openai:gpt-4", enable_scaffolds=True)

# Scaffold tracks file operations
agent.call("Read config.json")
agent.call("Write to output.txt")

# Access file history
files = agent.scaffolds["files"]
print(files.state.recent_files)
# ["config.json", "output.txt"]

# Scaffold provides context about files
agent.call("What files did I work with?")
# "You read config.json and wrote to output.txt"

Features:

File operation tracking
Working directory awareness
Recent file history
File relationship mapping

3. ShellScaffold

Command execution history and environment tracking:

agent = Agent(provider="openai:gpt-4", enable_scaffolds=True)

# Scaffold tracks commands
agent.call("Run: npm install")
agent.call("Execute: pytest")

# Access command history
shell = agent.scaffolds["shell"]
print(shell.state.command_history)
# [("npm install", "success"), ("pytest", "success")]

# Scaffold provides execution context
agent.call("What commands did I run?")
# "You ran npm install and pytest, both successful"

Features:

Command history tracking
Exit code monitoring
Environment variable awareness
Working directory tracking

Learn More

Complete guide to built-in scaffolds

Core Concepts

Scaffold State

Each scaffold maintains persistent state:

from egregore.core.context_scaffolds.base import BaseContextScaffold
from pydantic import BaseModel

class TaskState(BaseModel):
    tasks: list[str] = []
    completed: list[str] = []

class TaskScaffold(BaseContextScaffold):
    scaffold_type = "tasks"
    state_model = TaskState

    def render(self):
        # Render current state into context
        return TextContent(
            content=f"Active: {len(self.state.tasks)}, Completed: {len(self.state.completed)}",
            key="task_summary"
        )

State characteristics:

Type-safe with Pydantic models
Persistent across interactions
Accessible to other scaffolds
Survives agent restarts (if serialized)

Operations

Scaffolds expose operations as tools:

class TaskScaffold(BaseContextScaffold):
    scaffold_type = "tasks"

    @operation
    def add_task(self, task: str) -> str:
        """Add a new task."""
        self.state.tasks.append(task)
        return f"Added: {task}"

    @operation
    def complete_task(self, task: str) -> str:
        """Mark task as completed."""
        self.state.tasks.remove(task)
        self.state.completed.append(task)
        return f"Completed: {task}"

# Agent can use scaffold operations
agent = Agent(provider="openai:gpt-4", scaffolds=[TaskScaffold])
agent.call("Add task: write documentation")
# Scaffold's add_task() called automatically

Operation features:

Automatic tool generation
Type safety from annotations
Return values sent to agent
State changes trigger re-rendering

Learn More

Complete @operation decorator documentation

Reactive Rendering

Scaffolds automatically re-render when context changes:

class SmartScaffold(BaseContextScaffold):
    scaffold_type = "smart"
    _reactive = True  # Default: True

    def render(self):
        # Called automatically when context changes
        episode = self.agent.context.current_episode
        return TextContent(
            content=f"Current episode: {episode}",
            key="episode_tracker"
        )

    def should_rerender(self, ctx: ContextExecContext) -> bool:
        # Optional: control when to re-render
        return ctx.operation_type == "insert"  # Only on insertions

Reactive features:

Automatic re-rendering on context changes
Selective re-rendering via should_rerender()
Minimal performance overhead
All scaffolds reactive by default

Scaffold IPC

Scaffolds communicate through formal IPC:

class AnalyzerScaffold(BaseContextScaffold):
    scaffold_type = "analyzer"

    def render(self):
        # Read state from another scaffold
        notes = self.agent.scaffolds["notes"]
        note_count = len(notes.state.notes)

        # Write state for other scaffolds
        self.agent.state.set("analysis_status", "active", source="analyzer")

        return TextContent(
            content=f"Analyzed {note_count} notes",
            key="analysis"
        )

Learn More

Complete scaffold IPC documentation

Scaffold Lifecycle

Registration

Scaffolds are registered when creating an agent:

# Built-in scaffolds
agent = Agent(provider="openai:gpt-4", enable_scaffolds=True)

# Custom scaffolds
agent = Agent(
    provider="openai:gpt-4",
    scaffolds=[TaskScaffold, NotesScaffold]
)

# Access scaffolds
task_scaffold = agent.scaffolds["tasks"]
notes_scaffold = agent.scaffolds["notes"]

Initialization

Scaffolds initialize when first accessed:

class InitScaffold(BaseContextScaffold):
    scaffold_type = "init"

    def __init__(self, agent):
        super().__init__(agent)
        # Initialize state
        self.data = []
        # Setup connections
        self.setup_monitoring()

Rendering

Scaffolds render into context automatically:

class StatusScaffold(BaseContextScaffold):
    scaffold_type = "status"

    def render(self):
        # Return None to skip rendering
        if not self.should_render():
            return None

        # Return component to render
        return TextContent(
            content="Status: active",
            key="status",
            ttl=1,      # Expires after 1 turn
            cadence=1   # Re-renders each turn (sticky)
        )

Rendering triggers:

Initial agent creation
Context changes (if reactive)
Manual scaffold.render() call
State changes via operations

State Persistence

Scaffold state can be serialized:

import json

# Save scaffold state
state_data = agent.scaffolds["tasks"].state.model_dump()
with open("task_state.json", "w") as f:
    json.dump(state_data, f)

# Restore scaffold state
with open("task_state.json", "r") as f:
    state_data = json.load(f)

task_scaffold = agent.scaffolds["tasks"]
task_scaffold.state = TaskState(**state_data)

Use Cases

1. User Preferences

class PreferencesScaffold(BaseContextScaffold):
    scaffold_type = "preferences"

    class StateModel(BaseModel):
        theme: str = "light"
        language: str = "en"
        notifications: bool = True

    @operation
    def set_preference(self, key: str, value: str) -> str:
        """Update a preference."""
        setattr(self.state, key, value)
        return f"Set {key} to {value}"

    def render(self):
        prefs = [f"{k}: {v}" for k, v in self.state.dict().items()]
        return TextContent(
            content=f"Preferences:\n" + "\n".join(prefs),
            key="user_preferences"
        )

agent = Agent(provider="openai:gpt-4", scaffolds=[PreferencesScaffold])
agent.call("Set my theme to dark")
agent.call("Enable notifications")

2. Task Management

class TaskScaffold(BaseContextScaffold):
    scaffold_type = "tasks"

    class StateModel(BaseModel):
        tasks: list[dict] = []

    @operation
    def add_task(self, title: str, priority: str = "medium") -> str:
        """Add a task."""
        task = {"title": title, "priority": priority, "status": "pending"}
        self.state.tasks.append(task)
        return f"Added task: {title}"

    @operation
    def complete_task(self, title: str) -> str:
        """Complete a task."""
        for task in self.state.tasks:
            if task["title"] == title:
                task["status"] = "completed"
                return f"Completed: {title}"
        return f"Task not found: {title}"

    def render(self):
        pending = [t for t in self.state.tasks if t["status"] == "pending"]
        return TextContent(
            content=f"{len(pending)} pending tasks",
            key="task_count"
        )

3. Context Summarization

class SummaryScaffold(BaseContextScaffold):
    scaffold_type = "summary"

    def render(self):
        # Analyze recent messages
        thread = self.agent.thread.current
        message_count = len(thread.all_messages)

        # Track topics
        topics = self.extract_topics(thread)

        return TextContent(
            content=f"Conversation summary:\n"
                    f"Messages: {message_count}\n"
                    f"Topics: {', '.join(topics)}",
            key="conversation_summary",
            ttl=5  # Update every 5 turns
        )

    def extract_topics(self, thread):
        # Topic extraction logic
        return ["python", "documentation", "AI"]

4. Memory Consolidation

class MemoryScaffold(BaseContextScaffold):
    scaffold_type = "memory"

    class StateModel(BaseModel):
        short_term: list[str] = []
        long_term: list[str] = []

    def render(self):
        # Consolidate short-term to long-term
        if len(self.state.short_term) > 10:
            summary = self.consolidate(self.state.short_term)
            self.state.long_term.append(summary)
            self.state.short_term.clear()

        return TextContent(
            content=f"Memories: {len(self.state.long_term)} consolidated",
            key="memory_status"
        )

    def consolidate(self, memories: list[str]) -> str:
        # Use AI to summarize
        summary_agent = Agent(provider="openai:gpt-4")
        return summary_agent.call(
            f"Summarize these memories:\n" + "\n".join(memories)
        )

Best Practices

Use Pydantic for state

Type-safe state with validation:

from pydantic import BaseModel, Field

class StateModel(BaseModel):
    count: int = Field(ge=0)  # Non-negative
    items: list[str] = []
    enabled: bool = True

class MyScaffold(BaseContextScaffold):
    state_model = StateModel

Keep state focused

Each scaffold should have a single responsibility:

# Good: Focused scaffolds
class TaskScaffold(BaseContextScaffold): ...
class NotesScaffold(BaseContextScaffold): ...

# Bad: Kitchen sink scaffold
class EverythingScaffold(BaseContextScaffold):
    # tasks, notes, preferences, files, etc. - too much!

Use operations for mutations

Expose state changes as operations:

class ScaffoldWithOps(BaseContextScaffold):
    @operation
    def add_item(self, item: str) -> str:
        """Add item - exposed as tool."""
        self.state.items.append(item)
        return f"Added: {item}"

    # Don't: Modify state without @operation
    # Agent won't be able to use it

Control rendering frequency

Use TTL and selective re-rendering:

def render(self):
    return TextContent(
        content=self.generate_summary(),
        key="summary",
        ttl=10,     # Only re-render every 10 turns
        cadence=10  # Sticky behavior
    )

def should_rerender(self, ctx):
    # Only re-render on important changes
    return ctx.operation_type == "insert"

Performance Considerations

Rendering Overhead

Scaffolds add minimal overhead:

# Reactive scaffolds: ~5% overhead with 5+ scaffolds
agent = Agent(
    provider="openai:gpt-4",
    scaffolds=[S1, S2, S3, S4, S5]  # Minimal impact
)

# Disable reactivity if needed
class NonReactiveScaffold(BaseContextScaffold):
    _reactive = False  # No automatic re-rendering

State Size

Keep state manageable:

# Good: Bounded state
class BoundedScaffold(BaseContextScaffold):
    def render(self):
        # Keep only recent items
        if len(self.state.items) > 100:
            self.state.items = self.state.items[-100:]

# Bad: Unbounded growth
# State grows forever - memory issues

Operation Complexity

Keep operations fast:

# Good: Fast operations
@operation
def add_item(self, item: str) -> str:
    self.state.items.append(item)  # O(1)
    return "Added"

# Bad: Slow operations
@operation
def analyze_all(self) -> str:
    # Expensive AI call - blocks agent
    result = expensive_analysis(self.state.items)
    return result  # Consider async or background processing

What’s Next?

Creating Scaffolds

Learn how to build custom scaffolds

Scaffold IPC

Formal communication between scaffolds

Built-in Scaffolds

Complete guide to InternalNotes, FileManager, Shell

Retention Management

Managing scaffold state and persistence

Getting Started

Core Concepts

Scaffolds

Hooks

Workflows

Tools

Other Features

Guides - Getting Started

Guides - Basic Usage

Guides - Advanced

Guides - Testing & Debugging

Architecture

Contributing

Reference

Migration

API - Agent

API - Context

API - Scaffolds

API - Tools

API - Hooks

API - Providers

API - Workflow

API - Messaging

API - Accessors

API - Analytics

API - Configuration

Examples - Basic

Examples - Scaffolds

Examples - Workflows

Examples - Hooks

Examples - Advanced

Examples - Real-World

​Scaffolds Overview

​What are Scaffolds?

​The Scaffold Pattern

​Built-in Scaffolds

​1. InternalNotesScaffold

​2. FileManager

​3. ShellScaffold

Learn More

​Core Concepts

​Scaffold State

​Operations

Learn More

​Reactive Rendering

​Scaffold IPC

Learn More

​Scaffold Lifecycle

​Registration

​Initialization

​Rendering

​State Persistence

​Use Cases

​1. User Preferences

​2. Task Management

​3. Context Summarization

​4. Memory Consolidation

​Best Practices

​Performance Considerations

​Rendering Overhead

​State Size

​Operation Complexity

​What’s Next?

Creating Scaffolds

Scaffold IPC

Built-in Scaffolds

Retention Management

Scaffolds Overview

What are Scaffolds?

The Scaffold Pattern

Built-in Scaffolds

1. InternalNotesScaffold

2. FileManager

3. ShellScaffold

Core Concepts

Scaffold State

Operations

Reactive Rendering

Scaffold IPC

Scaffold Lifecycle

Registration

Initialization

Rendering

State Persistence

Use Cases

1. User Preferences

2. Task Management

3. Context Summarization

4. Memory Consolidation

Best Practices

Performance Considerations

Rendering Overhead

State Size

Operation Complexity

What’s Next?