You correct a caption on your phone. The picture is unchanged. Its location is unchanged. But an assistant has already used the old caption to produce a summary.
Which parts of its memory need to change?
That small edit exposes an important design problem for multimodal AI. An application needs to know which source version produced each derived result, what the result depends on and whether it remains suitable for the next answer. As more AI features arrive, that responsibility reaches beyond a single search index.
8DB puts a shared information layer at the center of that design. Its omnimodal architecture brings specialized representations together with native structures for confidence, provenance and governed storage. The opportunity for device applications is substantial: add new ways to understand information while keeping their evidence connected.
Give every view a source
A personal AI system might use image similarity to find a photo, a graph relationship to connect it to an event, and a time or location constraint to narrow the answer. Each method contributes something useful.
The shared record must also explain what those results mean. A coordinate captured by the device has a different origin from a location inferred from the image. A generated caption has a different status from words the user supplied. Keeping those distinctions available gives the application more control over which evidence it presents to a model.
In 8DB's projection architecture, specialized views belong to a common information model. The design lets graph, spatial, text and other operations address connected information while using the representations appropriate to their work.
Indexes still have to be built and maintained. The useful architectural aim is to share source identity, dependencies and rules for use across those operations.
One correction, a focused refresh
Consider a caption corrected from “Team dinner in Calgary” to “Client dinner in Edmonton.” An image-only embedding can remain useful because its input pixels did not change. A caption-based embedding needs refreshing. A summary that used the old caption needs reassessment.
The payoff is practical: a dependency-aware application can focus its work on affected results, preserve useful calculations and explain why an answer changed. The same structure can help a developer distinguish a corrected observation from a new interpretation.
Illustrated integration scenario · behavior to qualify
A caption changes. Which memories should follow?
Dependencies identify the work to revisit. The source image and its independent observations can remain useful.
Keep the unchanged inputs
Original photo bytes
Captured coordinates
Image-only embedding*
Refresh the affected views
Caption search index
Caption-based embedding
Summary using the old caption
Resolve the source revision, check the caller's current permission, and distinguish current evidence from results awaiting reassessment.
Build on the coordination that already exists
Device developers already have valuable tools for this job. Apple's Core Data integrates with Core Spotlight through persistent history, and its indexing notifications can identify the history token that an index has processed. SQLite's FTS5 external-content tables can refer back to source rows, with triggers maintaining the text index. Apple's integration walkthrough and SQLite's documentation explain these mechanisms.
8DB's proposed value reaches across a broader combination: media, relationships, similarity, location, confidence and governed evidence in one information layer. That makes the useful evaluation question concrete: how much application coordination is needed to keep this combination correct when information changes?
For a team adding several forms of AI retrieval, shared structures could reduce the number of separate conventions it must design and maintain. The benefit should be assessed on its actual workflow, including existing platform capabilities.
Start with a device path that already exists
BraidVault, our device application, gives this direction a concrete starting point. Its media test writes chunks through the engine, finalizes their media manifest, then requests a window around a playback position. It checks that the returned chunks match the original bytes. Protected payload storage and selective media retrieval meet in the same test.
Separate tests check that repeated content resolves to the same content address, that a large payload is stored through the engine's blob path, and that a different passphrase cannot recover a protected item. The project's June 22, 2026 report records these tests passing on an iPhone 15 Pro Max. These are device component results; they establish a useful starting point for a richer application.
For AI-derived information, 8DB's native insert and update messages can carry confidence, provenance and parent references. That gives the next layer a place to represent which observation contributed to a result. The governed key-derivation path combines a secret root with a defined governance context after a precheck. Each participating path still needs its configured checks and retention rules.
A team can therefore evaluate the architecture incrementally: start with protected content, connect derived records to it, then test the correction and permission behavior across the views its application actually uses. The caption-refresh scenario is the integration step to qualify. Our on-device retrieval article examines the complementary question of finding context that can support an answer.
Evaluate the memory your next feature will inherit
For a pilot, choose one real item and follow it through ingestion, search, correction and reopening. Check that each returned view identifies the source revision it describes. Change permission and verify what each caller can retrieve. Measure useful retrieval alongside the cost of keeping the views current.
That evaluation puts the architectural promise where it belongs: in an application people can use. A model can become better at seeing, searching and reasoning while the information beneath it keeps its relationships and history.
If you are building personal AI across photos, documents, maps or events, send us one workflow: what the user asks, which information supports the answer, and what must happen when that information changes. We can turn that into an evaluation plan covering useful answers, permitted access and the work needed to keep the evidence current.
Evidence scope Source-inspected device components · Proposed integration workflow. Explore the evidence. Methods and retained scope notes.
Continue the technical conversation
Where could this help your work?
Bring a research question, a database workload or an application you want to build. Let’s connect the ideas in this article to an evaluation that matters to your team.
Discuss this work