The technical and conceptual foundations of RAGMI — from the limitations of current AI agents to the Human Context Protocol designed to solve them.
This paper outlines a Human RAG (Retrieval Augmented Generation) solution that transforms human data into an active retrieval layer, allowing AI agents to better understand a human source without massive prompt overhead.
A major step forward in human understanding allows AI agents to shift from just being an assistant to a companion, opening the door to a range of previously unachievable scenarios.
Since the launch of OpenClaw there has been an explosion of people creating their own personal AI assistants. But a common complaint is that the relationship between the user and the agent is very much master and slave.
For many people, the concept of the perfect AI assistant is JARVIS — the fictional character from the Iron Man movies. JARVIS does what it's asked but has a relationship with Tony Stark that is as much companion as assistant. The question is: how do we enable our AI agents to know their users in the same way?
Getting the agent to ask the right questions is brittle and exhausting. The obvious solution is for the user to teach their agent all about themselves. LLMs use Retrieval Augmented Generation (RAG) to provide specialist knowledge — Human RAG provides the specialist knowledge of the person controlling the agent.
Each project within the RAGMI application is known as a folio — a container for all the media, stories, connections and events related to that subject. Current AI systems struggle with personal data because they lack a significance filter.
To an LLM, a receipt for a coffee and a receipt for a wedding ring are just two financial transactions. RAGMI solves this by introducing a processing pipeline that embeds human weighting directly into the data structure before the AI ever sees it.
To solve the bottleneck of agent context, RAGMI introduces the Human Context Protocol (HCP) — a dual-layer standard designed to bridge the gap between structured human history and active AI cognition.
The Factual Layer (JSON-LD) preserves the strict hierarchy, relationships, and weights of the user's life data. The Narrative Layer (Semantic Markdown) transforms that structured data into natural language optimised for RAG ingestion.
By unifying these layers, HCP creates a living digital legacy: a permanent archive for the user, and a preference-aware cognitive map for AI agents.
For older generations, Human RAG represents a closing window of opportunity. Media-linked autobiographical memories allow individuals to re-experience positive emotional states during periods of stress or decline — central to emotional regulation.
The most obvious application is for people suffering from dementia, where episodic memories can have a positive impact. A Human RAG-enabled agent can be instructed to surface specific episodic memories as an active wellbeing intervention.
RAGMI also allows individuals to capture not just facts, but perspective. After death, an agent trained on this data may act as an interactive memory store for descendants — disrupting the genealogy industry ($5–6.6B, projected to exceed $14–16B by the early 2030s).