# FERNme

A user-owned personalization memory layer for AI agents that models a person's preferences, habits and communication style as a fuzzy graph, with a deterministic zero-LLM core and optional human-approved enrichment.

FERNme builds a personalization model of each user as a sparse, fuzzily weighted graph node per site, updated with Hebbian-weighted relations, rather than by having a language model summarize the transcript on every turn. Its core writes and recalls memory deterministically with no model calls, and retrieval works by spreading activation across the graph; an optional, human-approved enrichment step can propose new connections using an agent or model the user already has, but nothing is written to memory until a person accepts it.

The prompt-facing summary given to an agent stores only deviations from a population prior, so it stays small and roughly flat in size as the underlying memory grows, and the whole model is inspectable, editable and deletable by the person it describes rather than acting as a hidden behavioral profile.

Website: https://www.fernme.dev