Patha: An AI Memory System

Patha is a novel AI memory system that moves beyond traditional storage and retrieval. It is built upon two core traditions: Vedic recitation, which utilized redundant encoding for preservation, and Australian Aboriginal songlines, where the landscape serves as an index for narrative traversal. This approach results in a system that separates retrieval from synthesis, offering significant token compression and efficient recall without relying on LLM tokens for synthesis queries.

Separation of Retrieval and Synthesis: Patha distinguishes between retrieving information and synthesizing answers, leading to a 6.5x compression of context compared to other systems.

Cognitive Status of Beliefs: Each belief is tagged with its origin (pramāṇa), mode of surfacing (vṛtti), and crystallization depth (saṃskāra → vāsanā), providing a structured understanding of knowledge.

Non-destructive Supersession: Old beliefs are not overwritten but marked as superseded, allowing users to trace the evolution of their thoughts.

Local-First and Interoperable: The belief store is a plain JSONL file, ensuring data ownership and compatibility with various MCP-compatible AI tools like Claude Desktop, Claude Code, and Cursor.

Performance Benchmarks: Patha demonstrates strong performance with a 1.000 R@5 on LongMemEval-KU and a 6.5x token reduction on multi-session tasks.

Patha offers multiple ways to interact, including as an MCP server for AI assistants, a command-line interface (CLI), and a Python library for developers building LLM applications. It also features a Streamlit viewer for visualizing the memory store.