EntoreX is a persistent, per-user context graph engine for consumer AI products. It stores what your users prefer, scores every result against those preferences, and evolves automatically from the choices they make. Drop it into any consumer product. Stateless AI answers the question. EntoreX answers your question.
You tell an AI assistant you're vegetarian, you hate chain hotels, you cap out at $300 per night — and next session it has forgotten all of it. Personalization today is a prompt you retype forever. Preferences aren't stated once. They're revealed — in what people pick, why they picked it, and what they rated well afterward.
EntoreX stores a typed graph per user: budget range, accommodation style, dietary needs, mobility needs, must-haves, avoids, interests — plus full trip history and every decision ever captured.
computeMatchScore() ranks any candidate 0–100 against the user's graph in real time, returning both the score and human-readable reasons and warnings. Explainable by construction — every score comes with its receipts.
Highly-rated stays reinforce style. Real spend tightens the budget band. Repeated reasons get promoted to must-haves. No manual tuning, no retraining, no prompt engineering. The graph gets sharper every time the user chooses something.
Every adjustment returns a reason string, so the score is always explainable. Budget fit adds 20 points when the price falls inside the user's min–max range. Over budget subtracts up to 20 points proportional to the overage. A style match adds 15. Each feature matching a must-have or interest adds 7, capped at 20. A walkable candidate adds 10 when walkable is a must-have. Dietary matches add 10. Each avoid hit subtracts 15 and stacks. A candidate similar to a past four-star-plus stay adds 5, plus 5 more when the price is within 20 percent.
Any stay rated four stars or higher adds its style to the user's preferred styles. The first trip sets the budget band to price times 0.7 through price times 1.3. Every trip after nudges the band via a 70/30 weighted average toward actual spend. Disliked attributes are appended to avoids.
EntoreX looks at the last 10 decisions. Any reason cited 3 or more times is auto-promoted to a must-have, which then earns +7 feature points on all future scoring. A user never fills in a preferences form: they book four hotels and mention walkable three times, and EntoreX promotes it automatically.
Pluggable persistence with zero logic changes to swap. ContextStore provides zero-dependency JSON file storage for local development. SqliteStore provides WAL-mode SQLite via better-sqlite3 for production. Reads hit an in-memory map cache first, then disk on miss; writes persist through on every mutation. The full API: getContext, getPreferences, getPastTrips, getDecisionHistory, getContextSummary, createUser, updatePreferences, addPastTrip, captureDecision, computeMatchScore.
EntoreX is domain-agnostic. Commerce: sizes, brands, materials, price ceiling, returns history. Food and dining: cuisines, allergies, spice tolerance, party size, favorite venues. Real estate: neighborhoods, commute limits, must-have features, dealbreakers. Media: genres, runtime tolerance, tone, watched-and-loved titles.
Vector databases retrieve similar documents. EntoreX stores structured, typed preferences and applies explicit scoring rules — every result comes with a human-readable reason, not a cosine distance.
Prompts don't evolve. EntoreX rewrites the user's profile from real behavior — budget bands shift, must-haves emerge — without anyone editing a prompt.
The opposite. computeMatchScore returns reasons and warnings alongside the number. Every point is attributable.
A new user scores at the baseline 50 and gets neutral ranking. The graph fills in from the first captured decision — no onboarding form required.
Wherever you point it. JSON on disk or SQLite locally; the StoreBackend interface takes any backend you implement.
No retraining. No prompt engineering. Just decisions. EntoreX is the memory layer for consumer AI — a consumer context graph engine that makes every AI interaction personal from the first session.