Agent Runtime
Intake routing, the responsible business agent, skills, profile context, memory recall, task ledger, and output rendering run on device.
NOW ON THE APP STORE · PRISMSAI CAMPUS OS
PRISMSAI is not a chat box that hands students over to a backend. Agents, memory, profiles, and task orchestration live on the phone. Only what the chosen model needs for the current inference travels through an isolated stateless relay.
PRISMSAI 1.0 · FreeiPhone + iPad · iOS 17+Available outside mainland China
Derived from traceable source history
Uncertainty remains explicit
A DIFFERENT AI ARCHITECTURE
Conventional AI products centralize conversation, memory, and orchestration in the cloud. PRISMSAI places the intelligent runtime on the iPhone and separates the model-content path from the management control plane.
Intake routing, the responsible business agent, skills, profile context, memory recall, task ledger, and output rendering run on device.
A separate process authenticates use, injects provider credentials, and forwards requests without a database, admin interface, or content logs.
The selected model reasons and returns a result. It does not own PRISMSAI's profile store, task ledger, personal knowledge, or agent orchestration.
Accounts, agent and skill versions, school knowledge retrieval, anonymous technical failures, and knowledge gaps use a separate control plane. Full conversations, profile text, memory bodies, and model answers do not enter it.
The boundary, precisely: stateless means content passes through memory without PRISMSAI persistence or management-plane visibility. It is not end-to-end encryption between the phone and model provider; the provider receives what it needs for inference. Data that must never leave the phone requires an on-device model.
WHAT WE BELIEVE
Student questions, personal profiles, and task orchestration stay on the student's device first.
Search both school and external knowledge. If reliable evidence is missing, make the gap visible.
From concise text to interactive HTML, present each result in the form that best fits the task.
PROVENANCE-FIRST MEMORY
PRISMSAI does not repeatedly stuff entire conversations into a model. It turns use into memory structures that are sourced, updateable, and deletable, so personalization can grow without losing its evidence boundary.
Preserve exact student inputs, selected images, and imported files under stable source IDs.
What actually happenedThe profile agent compresses durable preferences, goals, habits, and uncertainty into per-claim conclusions bound to source, freshness, and confidence.
What remains usefulA task receives only a minimal relevant profile. When more support is needed, the business agent retrieves detailed memory by locator.
Only what this task needsPreserves continuity with the responsible business agent without routing a follow-up through intake again.
Updates through real use and checks for age, conflict, weak evidence, and review needs.
Assesses semantic quality, source conflict, freshness, and confidence without inventing missing legacy citations.
THE PHONE IS THE TOOLBOX
With explicit user intent and iOS permission, the on-device runtime can read necessary context, prepare local actions, and return real system receipts.
Read nearby conflicts and create events or local notifications the student explicitly requests.
Use current coarse location without building location history, then hand off to Apple Maps when needed.
Process only content the user captures or selects, then add it to image memory and personal knowledge.
Read authorized step and sleep summaries without writing Health data or offering medical diagnosis.
Persist text, webpages, images, and files from other apps, then continue organization and indexing in the background.
Use system dictation and file selection in Unified Ask, then hand results to relevant applications.
THE AI CENTER FOUNDATION
The same on-device hub connects agents, memory, personal knowledge, device capabilities, and durable tasks. Each agent has a clear responsibility while sharing the same sources, identity context, and execution truth.
Intake understands the task; one business agent owns the result; memory, knowledge, and device tools participate as shared capabilities; an artifact agent is added only when presentation requires it.
Profile, intake, business, and artifact agents read the same stable source IDs, minimal profile, and task receipts instead of maintaining conflicting hidden memories.
The profile agent can ground a Today Tip in authorized memory, reminders, calendar, and device facts. It proposes; it does not act without confirmation.
Share Extension persists text, webpages, images, and files from other apps. Maps, calendar, and related tasks can be handed to the appropriate system app.
Future continuity requires explicit opt-in, revocation, auditability, and encryption. Today, personal memory, private knowledge, and task bodies remain authoritative on the current iPhone.
IMPLEMENTED, NOT CONCEPT ART
These are not roadmap slogans. Every capability maps to the current iOS app, agent contract, or durable runtime infrastructure.
Review supporting and conflicting sources, freshness, confidence, and next action without fabricating legacy source references.
Vision OCR and image classification combine with capture time, dimensions, and available location metadata to form local source capsules.
The iPhone chunks personal knowledge, combines lexical scoring with system semantic vectors, and audits index version, count, and content digest.
The runtime selects a correction skill by stage and failure code, asks the original agent to rebuild the complete object, and records every correction result.
Independent semantic verification checks claims against bound sources. Weak evidence produces a stable gap ID for the school's completion queue.
Tasks, events, confirmations, and corrections share one ledger that drives BGProcessing, Widget, Live Activity, notifications, and privacy-safe Spotlight summaries.
Infographics, mind maps, and flows render in an isolated WebView governed by CSP, navigation blocking, and evidence constraints.
Receipts capture stages, model attribution, knowledge queries, correction paths, and one runtime fingerprint; failures can flush anonymous diagnostics.
BUILDING IN PUBLIC · SINCE JULY 2026
Six weeks of architecture, product, knowledge, and live-runtime work—kept directly inspectable.
Loading updates…
QUESTIONS, ANSWERED
PRISMSAI 1.0 is free on the App Store for iPhone and iPad in supported storefronts outside mainland China. The Download button opens the official listing directly in the App Store app.
A local-first student AI campus assistant connecting questions, personal memory, knowledge retrieval, specialist agents, and task artifacts.
Planning, profiles, and task context stay on device first. The public website never serves as the student's private assistant client.
Course planning, campus policy and safety, clubs, schedules, reminders, knowledge organization, text, and interactive HTML artifacts.
STUDENT COMMONS
Share campus experience, discuss product ideas, and ask other students for help. Real names are not required. Do not post private information, grades, or account details.
Connecting to the community…
OPEN FEEDBACK
Bugs, missing campus knowledge, product ideas, and experience problems enter a real feedback queue. Contact information is optional.
Every submission has a receipt
Successful feedback returns a reference number you can keep.