Summary

Ori-Mnemos offers a local-first, persistent memory infrastructure for AI agents, prioritizing data sovereignty and eliminating cloud dependency. It provides engineers and power users with a high-performance, self-optimizing memory system that learns and adapts, ensuring full control over agent data.

How It Works

The system leverages a novel Recursive Memory Harness (RMH) framework, modeling human cognition on a knowledge graph. It employs ACT-R-inspired activation decay, spreading activation, Hebbian co-occurrence learning, and reinforcement learning on retrieval. Retrieval involves recursive graph traversal with sub-question decomposition, enabling the system to learn relevance, forget obsolescence, and optimize its retrieval pipeline. This approach delivers zero-infrastructure, high-performance retrieval directly on user data.

Quick Start & Requirements

• Install: npm install -g ori-memory
• Initialize: ori init my-agent
• Connect: Use ori bridge <client> --vault ~/brain for various MCP clients (e.g., claude-code, hermes, cursor).
• Prerequisites: Node.js/npm. Optional LLM integration requires compatible models.
• Docs: CLI commands and ori.config.yaml detail configuration.

Highlighted Details

• Benchmark Performance: Outperforms Mem0 on HotpotQA by 3.1x Recall@5 and 9.5x faster latency, using only Markdown and SQLite.
• Zero Cloud Dependency: Utilizes local embeddings (all-MiniLM-L6-v2) and SQLite, requiring no API keys for core functionality.
• Data Sovereignty: Stores memory as plain Markdown files, versioned with Git, ensuring portability and preventing vendor lock-in.
• Recursive Memory Harness (RMH): A novel framework treating memory as an evolving, navigable graph, incorporating complex retrieval and learning mechanisms.
• Cognitive Forgetting: Implements ACT-R-inspired decay and structural graph analysis, avoiding arbitrary TTLs.

Maintenance & Community

No specific details on maintainers, sponsorships, or community channels were found in the provided README. The project is at version v0.5.0.

Licensing & Compatibility

• License: Apache-2.0.
• Compatibility: Designed for local-first deployment with zero cloud dependencies. Data is stored in human-readable Markdown and versioned with Git, ensuring high portability and compatibility with standard developer tools. Suitable for commercial use.

Limitations & Caveats

The project is at version v0.5.0, indicating active development. While core functionality is heuristic-driven, optional LLM integration may be needed for advanced features. The novel RMH framework, while promising, represents a departure from traditional vector stores and may require further validation in diverse, long-term deployments.