Building continuity into AI-powered apps

Start with the experience, then design the context

A useful learning app needs continuity. The next piece of content should take account of what came before it. While building Aura — Self Refiner and My Exam Edge, I treated that as an application-design problem: what should the system remember, and what should it bring into the next generation request?

The inspiration was how experience informs our own thinking. A model call does not automatically inherit the history of previous calls. The application has to retain that history, select the relevant parts and supply them explicitly. That is the context-engineering work behind the interface.

Give recent content a place in the next request

Aura’s backend uses previously covered book titles to guide later content generation. In the optimized flow, a rolling window of up to 10 titles is passed into the prompt, which asks for a different book where possible. The generated result and updated history are stored together in an indexed, category-specific cache.

The March 2025 implementation maintains a shared content sequence. This helps vary the generated material across the sequence; personal memory for each reader would require a different state model.

My Exam Edge applies a related pattern to practice content. Its backend collects up to 60 recent cached question summaries or topics and supplies them alongside the syllabus domain. This gives the generator information about prior coverage instead of asking it to produce variety without context.

Separate content history from source knowledge

Remembering what was generated and knowing what an answer should be based on are different responsibilities. The current Exam Edge backend also implements retrieval-augmented generation, or RAG: selecting relevant passages from an ingested guide and supplying them with the generation request.

Guides are divided into sections, embedded and stored in Firestore with their examination and level. For a domain query, the retrieval path selects three nearby passages using vector similarity. Those passages become source material for the question and explanation. The July 2026 retrieval implementation adds source grounding to the earlier history-aware generation approach.

Own the system around the model

The architecture brings together context selection, structured output, caching and persistent state. Cached results can be served without another generation request; the stored history gives subsequent requests a basis for continuity. These are product and operating decisions as much as prompting decisions.

I approached this as an end-to-end product responsibility: define the experience, decide what state the application owns, and connect generation to a usable app. The model supplies content within that system; the application owns the continuity and delivery.

Design around the limits

History in a prompt encourages variety; it does not prove that repetition has been eliminated. Retrieved passages give the model reference material; they do not, by themselves, prove that every generated answer is correct.

The Exam Edge implementation can continue without guide grounding when no guide is available or retrieval fails. That fallback needs its own quality checks: an answer generated without a guide should be evaluated separately from one grounded in retrieved material. The design described here covers the backend implementations; the store listings describe the available app features.

Explore the apps

Aura and My Exam Edge are available on the App Store. Their listings describe supported devices and current product features.