Ori Mnemos ("Ori" for short) is an open-source, local-first persistent-memory infrastructure for AI agents. The name combines "Ori" (consciousness) with "Mnemos" (memory), and the project's motto is "Memory is sovereignty" — you own your memory. It was created by Aayo Awoyemi; the code is released under Apache-2.0, written in TypeScript and distributed as an npm package (ori-memory).
Memory is stored as plain Markdown files on disk, version-controlled with Git — with no database or cloud lock-in. Knowledge forms a graph: Markdown files with [[wiki-links]] act as directed edges. The system splits memory into three zones with distinct decay rates: Identity (self/), Knowledge (notes/) and Operations (ops/).
Retrieval combines four signals (four-signal retrieval fusion): semantic embeddings, BM25 keyword matching, Personalized PageRank and associative "warmth". Forgetting is based on base-level learning equations from ACT-R theory, and structurally critical nodes are protected using Tarjan's algorithm. The Recursive Memory Harness (RMH) performs graph-following retrieval, sub-question decomposition with convergence detection, and graph reshaping through Hebbian learning.
Ori exposes 16 tools over the Model Context Protocol (MCP) and integrates with clients such as Claude Code (full lifecycle integration with hooks), Hermes Agent, OpenCode, Cursor and Codex. On the HotpotQA benchmark the author reports, versus Mem0, 3.1× better recall, 2.1× higher F1 and 9.5× lower latency, running locally on Markdown + SQLite.
