Sugar

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

# 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):

claude mcp add sugar -- sugar mcp memory

Claude Code - Task server (optional):

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):

# Goose

goose configure

Select "Add Extension" -> "Command-line Extension"

Name: sugar

Command: sugar mcp memory

OpenCode - one command setup

sugar opencode setup

Skills

Sugar ships Agent Skills - folders of instructions that teach coding agents to apply Sugar's methodology. Skills live under skills/ and follow the Agent Skills specification (a SKILL.md with name/description frontmatter plus instructions).

Each skill includes an evaluation dataset (evals/evals.json) and is measured with NVIDIA SkillEvaluator against the OpenCode harness. See Skill Benchmarks for the Skill Lift results - sugar-memory and sugar-task-planner both show measured positive lift.

Global Memory

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

# 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:

Search strategy - project-first with reserved guideline slots:

1. Search the project store first (local context always wins)
2. Reserve slots for global guidelines (cross-project standards always surface)
3. Fill remaining slots with other global results
4. 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.

# 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:

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.

# 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:

/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.

# 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.

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:

Installation

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

pipx install sugarai

Upgrade / Uninstall:

pipx upgrade sugarai

pipx uninstall sugarai

Other installation methods

pip (requires venv activation each session)

pip install sugarai

uv

uv pip install sugarai

With semantic search (recommended for memory):

pipx install 'sugarai[memory]'

With GitHub integration:

pipx install 'sugarai[github]'

All features:

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

~/.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:

.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:

sugar:

dry_run: false loop_interval: 300 max_concurrent_work: 3

claude: enable_agents: true

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

Documentation

• Quick Start
• Memory System
• Skill Benchmarks
• CLI Reference
• Task Orchestration
• Goose Integration
• OpenCode Integration
• GitHub Integration
• Configuration Guide
• Troubleshooting

Requirements

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

Contributing

Contributions welcome. See CONTRIBUTING.md.

git clone ht