Observe
A consented event arrives from wherever you work — an IDE agent, a chat client, a website.
How FERNme Works
Each person is a sparse, fuzzily-weighted node in their own graph.
Edges move by arithmetic — the write path never calls a model.
Connect
A saturating Hebbian update strengthens co-occurring edges. No model call.
Decay
ACT-R decay weakens what stops being reinforced, so the graph can change its mind.
Recall
Spreading activation assembles a card of roughly 25 tokens, holding only deviations from the prior.
Learn
record_outcome() reinforces what worked and weakens what backfired.
Hebbian Learning
FERNme uses Hebbian principles: “neurons that fire together, wire together.”
The more concepts co-occur in your experience, the stronger their connection becomes.
Strong connection Weak connection
Fuzzy-Edge Network
Life is rarely black and white. FERNme uses fuzzy edges with strength scores on a 0–9 scale to represent how strongly two concepts are related.
Strong (7 – 9) Medium (4 – 6) Weak (0 – 3)
Person-Owned
Consent-gated throughout, with export and right-to-be-forgotten built into the engine.
One Profile, Many Agents
MCP and REST let a coding agent, a chat client, and a site share the same memory.
Deterministic by Default
Graph updates run without model calls on the write path, so behaviour stays predictable.
Open & Testable
Apache-2.0, 283 tests, and every benchmark reproducible from the repository.
FERNme is an Apache-2.0 research preview at v0.4, with 283 passing tests.
Published results are synthetic or LLM-authored; a real-human pilot is the pending next step.