Memoir is a high-performance semantic memory system designed for AI agents, offering a robust solution for AI memory management through Git-like version control. By replacing traditional opaque vector databases with transparent and versioned, cryptographically secure memory storage, Memoir facilitates efficient hierarchical semantic paths for organizing data.

The Importance of Versioned Memory for Agents

AI agents often face significant challenges such as context contamination, token rent, and memory drift. Memoir addresses these challenges by introducing version control to the agent's memory, thereby enhancing reliability and performance.

Key Challenges Addressed:

• Context Contamination: By eliminating the global variable anti-pattern, Memoir ensures that agents respect their Git state, thereby preventing undesirable interactions between experimental and stable code.
• Token Rent on Flat Files: Using global storage files can invalidate entire caches with minor updates, leading to inefficiencies. Memoir's design optimizes memory updates without incurring unnecessary costs.
• Memory as Code: Traditional memory management lacks version control, making it difficult to audit or revert erroneous data entries. Memoir provides commands such as memoir blame and memoir checkout to facilitate memory management with accountability.

Key Features

• Git-like Versioning: Offers branching, committing, merging, and rollback capabilities while ensuring cryptographic integrity of the memories.
• Semantic Paths: Utilizes meaningful paths instead of arbitrary UUIDs, making memory identification more intuitive (e.g., profile.professional.skills.python).
• Efficient Lookups: Achieves O(log n) lookup speed through hierarchical search mechanisms, significantly enhancing performance over costly vector operations.
• Memory Aggregation: Automatically consolidates related memories, maintaining coherence and reducing clutter.
• Clean Architecture: Separated layers for storage, classification, and search reduce complexity and improve maintainability.
• Multiple Search Engines: Select between keyword-based searches or intelligent searches powered by Language Models (LLMs).

Getting Started

Implementing Memoir is straightforward with a simple command-line interface and a streamlined installation process. Below are basic usage examples:

# Create a new memoir store
memoir new my-memoir-store
cd my-memoir-store

# Store a memory with a specified path
memoir remember "Sarah prefers tabs and 2-space indents" -p preferences.coding.style

# Retrieve a stored memory
memoir get preferences.coding.style

# Conduct a semantic search
memoir recall "what does Sarah prefer?"

# Launch the visual explorer
memoir ui

Explore the UI Gallery for a preview of the user interface.

Documentation and Resources

Comprehensive documentation is available at zhangfengcdt.github.io/memoir, covering:

• Quickstart for a fast overview of core functionalities.
• CLI Reference detailing every command and flag.
• API Reference for developers to integrate Memoir with other applications.

Contribution Opportunities

Memoir is currently in its alpha phase, and contributions are encouraged, especially from developers working with coding agents. Engage with the community by checking the issue tracker, submitting bug reports, or proposing enhancements.

Learn more about Memoir and its capabilities by visiting the project homepage.