# The Twin Is the Floor

A personal AI has to know the user's facts before it can help the user think.

EpisTwin matters because it describes a good floor for that problem. The floor is a user-rooted knowledge graph where personal data from calendars, photos, contacts, documents, and apps becomes explicit triples. The model lifts messy objects into structure, retrieves through graph topology, opens the raw object again when the graph loses detail, and gives deletion a specific target. Memory becomes an inspectable civil record. The statistical fog clears into objects, edges, and handles.

That is worth taking seriously.

The most important theft is the Information Object. A personal memory system should preserve the raw object, its origin, its timestamp, its permissions, its extracted entities, and the edges created from it. The graph then becomes a governed projection over evidence. The user can inspect a fact, see where it came from, correct it, redact it, or delete the underlying object. This is how memory earns trust at human scale.

The second theft is the role split: model as transducer, graph as source of truth. The model should read, summarize, extract, orchestrate, and ask for missing context. The graph should hold the facts that survive the moment. This keeps the system from treating every generated sentence as memory. It also makes correction possible, because correction attaches to a node, edge, source, or transform with locality and provenance.

The third theft is raw-object re-grounding. A graph is a compression surface. Compression creates power and loss at the same time. EpisTwin's strongest move is the admission that symbolic memory sometimes has to walk back to the original photo, document, screenshot, or page. The right memory system needs both the map and the preserved territory.

The fourth theft is the red pen. RUVA's best product instinct is that users are editors of their own lives. Inspectability, redaction, and precise forgetting are product features, trust features, and training-signal features simultaneously. Every correction teaches the system what the person means by accuracy, salience, privacy, and selfhood.

The fifth theft is benchmark discipline. PersonalQA-71-100 is useful because it tests cross-silo, temporal, multimodal personal reasoning. Hari needs a stronger benchmark in the same spirit: can the system reduce a person's epiplexity? Does it make the person's own pattern more legible to herself? Does it help the person publish the abstraction before the market compresses the insight into a generic feature?

The strategic mistake would be treating the twin as the destination. A personal knowledge graph answers what happened, where the evidence lives, and which fact can be corrected. A stance layer answers which memories matter, which corrections compound, what the person is becoming able to do, and which names should enter the open internet before the field settles around weaker language.

Hari's advantage lives in that layer.

EpisTwin gives the user a queryable self. Hari is building a queryable stance: a public graph that exposes how insight forms, how corrections change the map, how concepts earn names, how sources get routed into durable claims, and how a person plus agents can become a better abstraction engine over time. The personal twin is private infrastructure. The public radiant is cultural infrastructure.

This changes the competitive map.

The EpisTwin/RUVA authors are credible field shapers. The papers come from a Politecnico di Bari and Universita della Tuscia cluster with knowledge-graph, recommender-system, and applied-AI depth. RUVA has a demo surface. The Wideverse adjacency gives the group a plausible path from research artifact to practical product. The current evidence points to academic/demo-first motion, yet the ability to build is real.

The customer threat is low-to-medium today and rises if the demo becomes an app, SDK, or spin-off product line. The technical threat is medium because their primitives are exactly the primitives a trustworthy personal AI needs. The idea-market threat is higher. "EpisTwin" is a clean name attached to a March 2026 arXiv paper, and a separate April 2026 "Epistemic Twins" preprint already uses the phrase family for LLM knowledge auditing. The naming territory is moving.

The response is adoption plus asymmetry.

Adopt the memory floor. Use explicit objects, source pointers, graph extraction, redaction handles, raw-object re-grounding, and cross-silo tests. Keep vector search as an adapter. Keep selfhood in explicit objects. Let the personal graph become boring infrastructure as fast as possible.

Then publish the layer above it.

The open claim Hari should own: a twin records the person's state; a radiant changes the person's future abstraction capacity. The competitive moat lives in the correction loop, the naming loop, the provenance loop, the public graph, and the capacity to state the reason first. A competitor can copy a graph-backed memory UI. It is harder to copy a living public record of why the graph exists, what it is optimizing, how it corrects itself, and which insights it named before the field had settled.

That is the compounding move. Steal the red pen. Steal the raw-object handle. Steal the benchmark discipline. Steal the graph as civil record. Then make the twin ordinary by owning the meaning-layer above it.

## Sources

- EpisTwin arXiv: https://arxiv.org/abs/2603.06290
- RUVA arXiv: https://arxiv.org/abs/2602.15553
- RUVA demo: http://sisinf00.poliba.it/ruva/
- "Epistemic Twins" Zenodo preprint: https://zenodo.org/records/19910437
- Personal Knowledge Graphs research agenda: https://research.google/pubs/personal-knowledge-graphs-a-research-agenda/
- Personal Knowledge Graph ecosystem survey: https://arxiv.org/abs/2304.09572