Sugar

Persistent memory for AI coding agents.

Your AI agent starts every session with amnesia. The architecture decisions, conventions, and gotchas you explained last week are gone. Sugar is the local-first memory layer that remembers them for you - per project, across projects, on your machine.

Your memory. Your machine. Your data.

What Sugar Does

Sugar is a memory layer your AI coding agent can read and write directly:

Project memory - Decisions, preferences, error patterns, and research stored per-project
Global memory - Standards and guidelines shared across every project you work on
Semantic search - Retrieve relevant context by meaning, not just keywords
MCP integration - Your AI agent reads and writes memory directly during sessions
Local-first - SQLite on your disk, no API keys, fully offline-capable
Task queue - Optional autonomous execution, powered by the same memory layer

Quick Start

server.ts

# Install once, use in any project
pipx install sugarai

# Initialize in your project
cd ~/dev/my-app
sugar init

# Store what you know
sugar remember "We use async/await everywhere, never callbacks" --type preference
sugar remember "JWT tokens use RS256, expire in 15 min - see auth/tokens.py" --type decision
sugar remember "When tests fail with import errors, check __init__.py exports first" --type error_pattern

# Retrieve it later
sugar recall "authentication"
sugar recall "how do we handle async"

Your AI agent can also read and write memory directly - no copy-pasting required.

MCP Integration

Connect Sugar's memory to your AI agent so it can access project context automatically.

Claude Code - Memory server (primary):

Terminal

claude mcp add sugar -- sugar mcp memory

Claude Code - Task server (optional):

Terminal

claude mcp add sugar-tasks -- sugar mcp tasks

Once connected, Claude can call store_learning to save context mid-session and search_memories to pull relevant knowledge before starting work. The memory server works from any directory - global memory is always available even outside a Sugar project.

Other MCP clients (Goose, Claude Desktop):

bash

# Goose
goose configure
# Select "Add Extension" -> "Command-line Extension"
# Name: sugar
# Command: sugar mcp memory

# OpenCode - one command setup
sugar opencode setup

Global Memory

Some knowledge belongs to you, not just one project. Coding standards, preferred patterns, security practices - these should follow you everywhere.

bash

# Store a guideline that applies to all your projects
sugar remember "Always validate and sanitize user input before any DB query" \
  --type guideline --global

sugar remember "Use conventional commits: feat/fix/chore/docs/test" \
  --type guideline --global

# View your global guidelines
sugar recall "security" --global
sugar memories --global

# Search works project-first, but guidelines always surface
sugar recall "database queries"
# Returns: project-specific memories + relevant global guidelines

Global memory lives at ~/.sugar/memory.db. Project memory lives at .sugar/memory.db. When you search, project context wins - but guideline type memories from global always appear in results so your standards stay visible.

Via MCP, pass scope: "global" to store_learning to save cross-project knowledge directly from your AI session.

Memory types: decision, preference, file_context, error_pattern, research, outcome, guideline

Full docs: Memory System Guide

How Memory Works

Sugar uses two SQLite databases and a tiered search strategy.

Two stores:

Project store (.sugar/memory.db) - context specific to one project
Global store (~/.sugar/memory.db) - knowledge that applies everywhere

Seven memory types, each with different retrieval behavior:

Type	Purpose	TTL
`decision`	Architecture and implementation choices	Never
`preference`	How you like things done	Never
`file_context`	What files and modules do	Never
`error_pattern`	Bugs and their fixes	90 days
`research`	API docs, library findings	60 days
`outcome`	What worked, what didn't	30 days
`guideline`	Cross-project standards and best practices	Never

Search strategy - project-first with reserved guideline slots:

Search the project store first (local context always wins)
Reserve slots for global guidelines (cross-project standards always surface)
Fill remaining slots with other global results
Deduplicate across both stores

This means a mature project's local context dominates results. A new project with no local memory gets global knowledge automatically. And your guidelines are always visible regardless.

Search engine: Semantic search via sentence-transformers (all-MiniLM-L6-v2, 384-dim vectors) with sqlite-vec. Falls back to SQLite FTS5 keyword search, then LIKE queries. No external API calls - everything runs locally.

bash

# Install with semantic search (recommended)
pipx install 'sugarai[memory]'

# Works without it too - just uses keyword matching
pipx install sugarai

MCP tools available to your AI agent:

Tool	What it does
`search_memory`	Search both stores, returns results with scope labels
`store_learning`	Save a memory (pass `scope: "global"` for cross-project)
`recall`	Get formatted markdown context for a topic
`get_project_context`	Full project summary including global guidelines
`list_recent_memories`	Browse recent memories by type

MCP resources:

sugar://project/context - project summary
sugar://preferences - coding preferences
sugar://global/guidelines - cross-project standards

Task Queue

The task queue lets you hand off work and let it run autonomously. It reads from the same memory store, so Sugar already knows your preferences and patterns before it starts.

bash

# Add tasks
sugar add "Fix authentication timeout" --type bug_fix --urgent
sugar add "Add user profile settings" --type feature

# Start the autonomous loop
sugar run

Sugar picks up tasks, executes them with your configured AI agent, runs tests, commits working code, and moves to the next task. It runs until the queue is empty or you stop it.

Delegate from Claude Code mid-session:

Code

/sugar-task "Fix login timeout" --type bug_fix --urgent

Advanced task options: New in 3.10: Task Orchestration decomposes large features into a 4-stage workflow (research, plan, implement, review) with specialist agent routing and dependency-ordered sub-tasks.

bash

# Orchestrated execution - 4-stage workflow (New in 3.10)
sugar add "Add OAuth authentication" --type feature --orchestrate

# Iterative mode - loops until tests pass
sugar add "Implement rate limiting" --ralph --max-iterations 10

# Check queue status
sugar list
sugar status

Full docs: Task Orchestration

Autonomous Issue Resolution (optional)

Because Sugar remembers your codebase and conventions, it can also resolve routine issues autonomously. Point it at a GitHub repo, configure which labels to act on (security, bug, dependabot), and Sugar will read each issue, implement the fix, run your tests, and open a PR.

Code

Labeled issue appears on GitHub
  -> Sugar picks it up (label filter: "security", "dependabot", "bug")
  -> AI agent reads the issue, analyzes the affected code
  -> Fix implemented, tests run locally
  -> PR opened - you review and merge

This is one application of the memory layer, not the headline. Use Sugar purely as memory, or enable resolution - your choice. See workflow examples for security auto-fix, bug triage, test coverage, and more.

Supported AI Tools

Works with any CLI-based AI coding agent:

Agent	Memory MCP	Task MCP	Notes
Claude Code	Yes	Yes	Full support
OpenCode	Yes	Yes	`sugar opencode setup`
Goose	Yes	Yes	Via MCP
Aider	Via CLI	Via CLI	Manual recall

Installation

Recommended: pipx - installs once, available everywhere, no venv conflicts:

bash

pipx install sugarai

Upgrade / Uninstall:

bash

pipx upgrade sugarai
pipx uninstall sugarai

Other installation methods

pip (requires venv activation each session)

Terminal

pip install sugarai

bash

uv pip install sugarai

With semantic search (recommended for memory):

bash

pipx install 'sugarai[memory]'

With GitHub integration:

bash

pipx install 'sugarai[github]'

All features:

bash

pipx install 'sugarai[all]'

Sugar is project-local by default. Each project gets its own .sugar/ folder with its own database and config. Global memory lives at ~/.sugar/. Like git - one installation, per-project state.

Project Structure

Code

~/.sugar/
└── memory.db          # Global memory (guidelines, cross-project knowledge)

~/dev/my-app/
├── .sugar/
│   ├── sugar.db       # Project memory + task queue
│   ├── config.yaml    # Project settings
│   └── prompts/       # Custom agent prompts
└── src/

Recommended .gitignore:

gitignore

.sugar/sugar.db
.sugar/sugar.log
.sugar/*.db-*

Commit .sugar/config.yaml and .sugar/prompts/ to share settings with your team.

Configuration

.sugar/config.yaml is created on sugar init:

yaml

sugar:
  dry_run: false
  loop_interval: 300
  max_concurrent_work: 3

claude:
  enable_agents: true

discovery:
  github:
    enabled: true
    repo: "user/repository"

Documentation

Requirements

Python 3.11+
A CLI-based AI agent: Claude Code, OpenCode, Aider, or similar

Contributing

Contributions welcome. See CONTRIBUTING.md.

bash

git clone https://github.com/roboticforce/sugar.git
cd sugar
uv pip install -e ".[dev,test,github]"
pytest tests/ -v

License

Dual License: AGPL-3.0 + Commercial

Open Source (AGPL-3.0): Free for open source and personal use
Commercial License: For proprietary use - sugar.roboticforce.io/licensing

Sugar is provided "AS IS" without warranty. Review all AI-generated code before use.