mnemosyne

Mnemosyne provides a robust, local-first memory system designed specifically for AI agents, consolidating learned information in a single SQLite file. This innovative solution serves agents efficiently while adhering to the principles of auditable memory, ensuring that each piece of information is accompanied by its source and associated trust level. With Mnemosyne, agents start with the most reliable, current evidence, enhancing decision-making capabilities.

Key Features:

• Local-First Storage: Mnemosyne operates offline and does not require an external database server or Docker setup, making it an accessible tool for various AI projects.
• Evidence-Based Memory: Each entry in the memory database retains metadata that relates to its source, including trust tiers, ensuring accuracy and reliability throughout operations.
• Dynamic Corrections: When a correction is made, any associated dependencies are automatically updated, ensuring that no outdated lessons are repeated by the agents.
• Audit Capabilities: The system records past beliefs, allowing users to replay decisions at any time, providing transparency in the decision-making process.
• Versatile Integration: Mnemosyne supports multiple access methods: MCP for frameworks like Claude Code, Cursor, and Codex; along with TypeScript SDK and CLI, facilitating easy integration into various development environments.

Proven Performance:

Mnemosyne boasts impressive benchmark results, achieving a score of 409 out of 500 in the LongMemEval-S, illustrating its capacity for effective memory management. The innovative functionalities in the latest releases include inferred dependency tracking, which aids in maintaining continuity across updates.

Installation and Getting Started:

To incorporate Mnemosyne in a project, install it via npm, either the stable version or the pre-release for cutting-edge features. For instance:

npm install mnemosy-ai # Stable version
npm install mnemosy-ai@next # Pre-release

To run a demonstration, use the npx command, which showcases the integral features of the memory system:

npx -y -p mnemosy-ai demo

Example Usage:

Here is a simple code snippet illustrating how to set up a memory instance with Mnemosyne, correct a piece of memory, and retrieve its state:

import { createLocalMemory } from 'mnemosy-ai/local';

const memory = createLocalMemory({
  path: './agent-memory.sqlite',
  workspaceId: 'catalogue',
  agentId: 'designer',
});

const brief = memory.store({
  text: 'Product images must be 1200 by 900 pixels.',
  kind: 'fact',
  key: 'catalogue.image-size',
  source: { uri: 'brief://catalogue/v1' },
  trust: 'observed',
  visibility: 'workspace',
});

memory.correct(brief.id, {
  text: 'Product images must be 1600 by 1200 pixels.',
  source: { uri: 'brief://catalogue/v2' },
  reason: 'The delivery specification changed.',
});

console.log(memory.get(brief.id)?.status); // Outputs: invalidated

For more detailed insights into capabilities, corrections, or other advanced features, refer to the comprehensive documentation and begin constructing powerful AI agents with Mnemosyne.