SAM Features

April 16, 2026 · View on GitHub

Complete feature reference for SAM (Synthetic Autonomic Mind)


Overview

SAM is a native macOS AI assistant with a comprehensive set of features for conversation, research, document management, file operations, image generation, math, and voice control. This document covers every feature in detail.


Conversations

Unlimited Conversations

Create as many conversations as you want. Each one is saved automatically to ~/Library/Application Support/SAM/conversations/ and persists across app restarts.

Automatic Saving

Every message is saved as it's sent. You never need to manually save. If SAM crashes or you force-quit, your conversations are preserved.

Conversation Export

Export any conversation in two formats:

  • JSON - Full structured export with metadata, timestamps, and tool calls
  • Markdown - Human-readable format suitable for sharing or archiving

Conversation Import

Import previously exported conversations back into SAM.

Context Isolation

Conversations are isolated by default. The AI in one conversation has no knowledge of other conversations. This is a privacy feature - you can have a work conversation and a personal conversation without cross-contamination.

To share context between conversations, use Shared Topics.

Model Switching

Switch AI models at any point during a conversation. The full conversation history carries forward to the new model. Useful for starting with a fast model and switching to a more capable one for complex tasks.

Per-Conversation Settings

Each conversation stores its own UI state:

  • Panel visibility (tool cards, performance, etc.)
  • Scroll position
  • Active model

AI Provider Support

Cloud Providers

OpenAI

  • Models: GPT-4o, GPT-4, GPT-3.5 Turbo, o1, o3
  • Streaming support with real-time token delivery
  • Function calling and tool use

GitHub Copilot

  • Models: GPT-4o, Claude 3.5, o1 (varies by subscription)
  • Device flow authentication (no manual API key needed)
  • Automatic token refresh

DeepSeek

  • Models: DeepSeek Chat, DeepSeek Coder
  • Cost-effective alternative for coding and general tasks

Google Gemini

  • Models: Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.0 Flash
  • Large context windows (up to 1M tokens)
  • Strong multimodal capabilities

MiniMax

  • Models: MiniMax-M2.7, MiniMax-M2.5, high-speed variants
  • 128K token context window
  • Competitive pricing with good tool use

OpenRouter

  • Access 100+ models from multiple providers through a single API
  • Automatic model routing and load balancing

Note on Claude models: SAM does not include a direct Anthropic provider. Claude-family models are available through OpenRouter.

Local Models

MLX (Apple Silicon)

  • Runs models using Apple's MLX framework
  • Metal GPU acceleration for fast inference
  • Efficient memory usage with unified memory architecture
  • Support for Hugging Face model format
  • Model caching and lazy loading
  • LRU eviction for memory management

llama.cpp (Any Mac)

  • Runs GGUF-format models on Apple Silicon or Intel
  • Compiled as a native framework (included as git submodule)
  • CPU and GPU inference modes

Model Download Manager

  • Browse available models from Hugging Face
  • Download progress tracking with pause/resume
  • Model size estimates before download
  • Automatic storage management

Custom Endpoints

Connect to any OpenAI-compatible API:

  • Local servers (Ollama, LM Studio, text-generation-webui)
  • Self-hosted inference servers
  • Custom routing proxies

Autonomous Tool System

SAM has 8 consolidated tools that the AI uses autonomously to accomplish tasks. Each consolidated tool contains multiple operations.

File Operations

OperationWhat It Does
read_fileRead file content with optional line range
list_dirList directory contents (recursive or flat)
get_errorsCheck a file for compilation/lint errors
get_file_infoGet file metadata (size, type, modified time)
read_tool_resultRead large results in chunks
file_searchFind files matching a glob pattern
grep_searchSearch file contents with text or regex
semantic_searchFind files by meaning using NLP
list_usagesFind all references to a symbol
create_fileCreate a new file with content
replace_stringFind and replace text in a file
multi_replace_stringBatch replacements across files
insert_editInsert content at a specific location
rename_fileRename or move a file
delete_fileDelete a file or directory
create_directoryCreate a directory (with parents)

Path Authorization:

  • Inside ~/SAM/ working directory: auto-approved
  • Outside working directory: requires user confirmation

Web Operations

OperationWhat It Does
researchMulti-source web research with automatic memory storage
retrieveAccess previously stored research
web_searchSearch the web (Google, Bing, DuckDuckGo)
serpapiDirect SerpAPI access (Google, Bing, Amazon, eBay, TripAdvisor, Walmart, Yelp)
scrapeFull WebKit page rendering with JavaScript support
fetchFast HTTP fetch without JavaScript

Research is SAM's most powerful web tool. It conducts multi-source research, synthesizes findings, and stores results in conversation memory for future reference.

Document Operations

OperationWhat It Does
document_importImport PDF, DOCX, XLSX, or TXT files into the conversation
document_createGenerate PDF, DOCX, PPTX, TXT, or Markdown files
get_doc_infoGet document metadata

Imported documents are chunked and indexed with vector embeddings for semantic search.

Memory Operations

OperationWhat It Does
search_memorySemantic search across conversation memories
store_memorySave important information for later
list_collectionsList available memory collections
recall_historyRecall conversation history by topic

Todo Operations

OperationWhat It Does
readGet current task list
writeCreate or replace task list
updateUpdate task status
addAdd new tasks

The AI uses todos to track multi-step work. It creates a plan, marks tasks in-progress as it works, and marks them complete as it finishes.

Math Operations

OperationWhat It Does
calculateEvaluate mathematical expressions via Python
computeRun arbitrary Python code for complex calculations
convertUnit conversions (temperature, length, weight, volume, speed, data, time)
formulaNamed financial and practical formulas

Available Formulas: tip, mortgage, bmi, compound_interest, percentage, markup, discount, area_circle, area_rectangle, volume_cylinder, speed_distance_time, sales_tax, gpa, fuel_cost, cooking, retirement, debt_payoff, debt_strategy, budget, loan_comparison, savings_goal, net_worth, paycheck, and inflation.

All math is computed by a real Python 3 interpreter - no AI approximation.

Image Generation

OperationWhat It Does
generateGenerate images from text descriptions

Connects to a remote ALICE server for GPU-accelerated Stable Diffusion image generation. Supports multiple models, automatic model discovery, and server health monitoring.

User Collaboration

OperationWhat It Does
request_inputPause and ask the user a question

Used by the AI when it needs clarification, confirmation for destructive operations, or user decisions.


Memory System

Vector RAG (Retrieval-Augmented Generation)

SAM uses vector embeddings to index conversations and documents for semantic search. This means you can find information by meaning, not just keywords.

How it works:

  1. Text is split into chunks
  2. Each chunk gets a vector embedding via Apple's NaturalLanguage framework
  3. Embeddings are stored in a per-conversation SQLite database
  4. Queries are embedded and matched against stored vectors
  5. Top matches are returned with similarity scores

Search across all conversations simultaneously. The search interface uses the same semantic matching to find relevant discussions regardless of exact wording.

Import documents into any conversation. The content is chunked, embedded, and stored for semantic retrieval. You can then ask questions about the document naturally.

Supported Formats: PDF, DOCX, XLSX, TXT, MD, CSV

Context Window Management

SAM uses context archival to manage long conversations:

  • Messages are archived to a per-conversation SQLite database when the context window fills
  • Archived chunks include summaries, key topics, and timestamps
  • Semantic search retrieves relevant archived context automatically
  • Pinned messages are preserved and always included in context
  • Dynamic context sizing based on model limits

Voice Control

Wake Word Detection

Say "Hey SAM" to activate voice input. Uses Apple's on-device speech recognition for privacy.

Speech-to-Text

Real-time speech transcription using Apple's Speech framework. All processing happens on-device.

Text-to-Speech

SAM speaks responses aloud using macOS native voices:

  • Streaming TTS - Starts speaking as soon as the first sentence is ready
  • Markdown stripping - Removes formatting before speaking
  • Voice selection - Choose from any installed macOS voice
  • Speed control - Adjustable rate from 0.5x to 1.5x

Audio Device Management

Select specific input (microphone) and output (speaker) devices. SAM detects all connected audio devices and updates the list when devices change.


Image Generation (ALICE)

Connect to a remote ALICE server for Stable Diffusion image generation:

  • Automatic model discovery - Detects models loaded on the ALICE server
  • Multiple model support - SD 1.5, SDXL, and custom models
  • Health monitoring - Connection status and server availability
  • No local GPU required - Generation happens on the ALICE server
  • In-chat display - Generated images appear directly in the conversation

SAM-Web (Remote Access)

Access SAM from any device on your local network through a web browser.

Features:

  • Full chat interface
  • Mini-prompts
  • Model selection
  • Conversation management
  • Responsive design (desktop, tablet, mobile)
  • Token-based authentication

Requirements:

  • SAM running on your Mac with API server enabled
  • SAM-Web deployed
  • Both devices on the same network

Shared Topics

Cross-conversation workspaces backed by SQLite:

  • Named topics - Create topics for projects, research areas, or any ongoing work
  • Shared workspace - All conversations in a topic share a file directory
  • Shared entries - Store and retrieve data accessible to any conversation in the topic
  • Optimistic locking - Safe concurrent access from multiple conversations
  • Audit trail - All operations are logged

System Prompt Customization

Built-In Personalities

SAM includes personality configurations that affect communication style:

  • Tone, vocabulary, and response style
  • Configurable personality traits
  • Per-conversation or global settings

Custom System Prompts

Edit the system prompt that defines SAM's behavior:

  • View and modify the active system prompt
  • Save custom system prompt templates
  • Switch between configurations
  • Component library for building modular prompts

Mini-Prompts

Quick-action templates for common tasks. Mini-prompts provide pre-configured instructions that you can invoke with a click instead of typing out full instructions each time.


API Server

SAM runs a local HTTP server (Vapor-based) that provides:

  • OpenAI-compatible API - Connect external tools that speak the OpenAI protocol
  • SAM-Web interface - Web-based access from other devices
  • Authentication - Token-based auth for all API endpoints
  • SSE streaming - Real-time streaming responses
  • CORS support - Configurable cross-origin access

The server runs on port 8080 by default and is disabled until explicitly enabled in Settings.


Performance Monitoring

Real-time metrics visible in the app:

  • Memory (RSS) - Application memory usage
  • Inference speed - Tokens per second for local models
  • Context usage - Current token count vs. model limit
  • API latency - Response time for cloud providers

Auto-Updates (Sparkle)

SAM uses the Sparkle framework for automatic updates:

  • Stable channel - Production releases (default)
  • Development channel - Pre-release builds (opt-in in Settings)
  • Background checks - SAM checks for updates periodically
  • Secure updates - Code-signed and notarized builds
  • Appcast feeds - Separate feeds for stable and development channels

Mermaid Diagrams

SAM renders Mermaid diagrams inline in conversations. When the AI generates a Mermaid code block, SAM renders it as a visual diagram. Supports 15 diagram types including flowcharts, sequence diagrams, class diagrams, state diagrams, and more.


Think Tags

When supported models use extended thinking (e.g., Claude's thinking blocks), SAM displays the reasoning process in collapsible sections. You can see the AI's step-by-step reasoning without it cluttering the response.


Logging

SAM uses structured logging via swift-log:

  • Log levels: debug, info, warning, error
  • Logger labels follow com.sam.<module> convention
  • Logs accessible via Help > Show Logs
  • Useful for troubleshooting and bug reports

Keyboard Shortcuts

ShortcutAction
⌘NNew conversation
⌘KClear current conversation
⇧⌘RRename conversation
⇧⌘DDuplicate conversation
⇧⌘EExport conversation
⌘⌫Delete conversation
⌘FSearch conversations
⇧⌘/Show help
⌘,Open Settings
⌘WClose window
EnterSend message
Shift+EnterNew line in message

Version Scheme

SAM uses date-based versioning:

  • Stable: YYYYMMDD.RELEASE (e.g., 20260110.1)
  • Development: YYYYMMDD.RELEASE-dev.BUILD (e.g., 20260110.1-dev.3)

See VERSIONING.md for details.