aayoawoyemi/Ori-Mnemos

What it solves

Ori Mnemos provides a persistent, local-first memory infrastructure for AI agents. It solves the problem of context window limitations and "forgetting" across sessions by treating memory as a navigable knowledge graph rather than a flat list of documents, allowing agents to maintain a consistent identity and retrieve complex information without relying on expensive cloud databases.

How it works

Ori implements a "Recursive Memory Harness" (RMH) that treats memory as an environment to be navigated. It uses markdown files on disk as the primary storage, with wiki-links acting as edges in a knowledge graph. Retrieval is performed through a four-signal fusion of semantic embeddings, BM25 keyword matching, Personalized PageRank, and "associative warmth" (based on ACT-R cognitive decay models).

To improve over time, the system employs three layers of retrieval intelligence: Q-value reranking based on session outcomes, Hebbian learning to create co-occurrence edges between notes retrieved together, and a contextual bandit meta-learner that optimizes which retrieval stages to run for specific query types.

Who it’s for

It is designed for developers building AI agents that require long-term memory, persistent identity across different clients (like Claude Code, Cursor, or Hermes), and full ownership of their data without vendor lock-in.

Highlights

• Local-First Architecture: Uses markdown files and SQLite, requiring no cloud APIs or external database infrastructure.
• Cognitive Forgetting: Implements ACT-R base-level learning equations to decay notes based on relevance and usage.
• Recursive Exploration: Decomposes complex questions into sub-questions and traverses the graph until convergence is reached.
• MCP Integration: Provides a Model Context Protocol (MCP) server with 16 tools for agents to manage their own memory.
• High Performance: Outperforms incumbents like Mem0 on multi-hop retrieval benchmarks (HotpotQA) with significantly lower latency.