Distilling a person: high-fidelity computational identity from the trace of a life

The naive approach to a digital twin —index the full history in a vector store and retrieve by similarity— fails for structural, not scale, reasons. Retrieval by semantic proximity does not preserve temporal order, cannot separate interlocutors that share lexical context, and favors literal repetition over stylistic generalization. In memory terms, it yields a purely associative system with no timeline and no model of the people involved: it recalls fragments, it does not reconstruct a life.

The working hypothesis is that value lies in compression, not accumulation: the history is raw material, not product. To distill is to infer who the person is —their idiolect, their bonds, their biography and their values— from the evidence, and to represent that compactly, queryably and auditably. Machine learning discovers the corpus's latent structure; the language model intervenes only at the close, over already-organized aggregates. The approach inherits a concrete precedent from the same lab: Aria, a persistent-identity entity built on curated memory and retrieval by meaning (see the Engram Project study); the digital twin takes that architecture from manual curation to the automatic distillation of a real person's history.

The architecture does not imitate the brain by decorative analogy; it borrows principles of memory organization because they resolve exactly the failures of the dump approach. Since Tulving, the psychology of memory distinguishes episodic memory —events situated in time and place— from semantic memory —general knowledge detached from its origin [1]—. A twin needs both: retrievable episodes (what happened, with whom, when) and consolidated knowledge (who this person is, in general). Systems neuroscience describes how they interlock: the complementary learning systems theory holds that the hippocampus rapidly encodes specific episodes while the neocortex slowly integrates that knowledge, through consolidation, into stable and generalizable representations [3, 4]. The physical trace of a memory —the engram— is today an empirical object of study, not a metaphor [2]; the name of this lab's sibling program is no accident.

The engine mirrors that division of labor. The indexed corpus and temporal graph play the role of the episodic-relational system of fast retrieval; identity synthesis and, above all, the style adapter trained into the weights play the role of semantic consolidation: the idiolect stops being retrieved as an example and becomes a disposition of the model, just as a motor skill stops depending on the hippocampus once consolidated in the cortex.

On memory rests identity. The self-memory system model describes the self as a construction sustained by autobiographical memory, hierarchically organized by lifetime periods and themes [5]; narrative psychology adds that adult identity is, in large part, the internalized story a person tells about themselves [6]. Hence the goal of distillation is not a message dump but structured artifacts —biography, bonds, values, voice— that reconstruct that story. Two psychometric anchors close the frame: stable traits are modeled with the Five-Factor Model (OCEAN), preferred for its validity over typologies with no empirical basis [7], on the grounds that language use is a measurable individual marker of a person's psychology [8]; and the value system with Schwartz's theory of basic values [9, 10].

The field organizes into three fronts. In twin/persona architectures, Second Me formalizes a layered pattern —raw data, summaries and adapted weights (PEFT)— that matches this engine's almost point for point and validates the design in practice [18]. In agent memory, generative agents introduce a memory stream with asynchronous reflection that consolidates recent experiences into higher-level beliefs [16], and MemGPT treats the context window as a managed memory hierarchy [17]; both patterns reappear here as the background consolidation thread.

In style and persona modeling, stylometry and authorship verification provide objective measures of “who wrote this” [11], which this work reuses as a fidelity metric. The machine-learning components are standard and deliberately reproducible: multilingual sentence embeddings [12, 13], neural topic modeling with BERTopic [14], and low-rank adaptation (LoRA) to fix style without retraining the full model [15]. The contribution is not inventing components but articulating them under two constraints the market does not combine: data sovereignty and measured fidelity.

The pipeline: ⓪ mode-configurable privacy → ① parse and normalize with turn fusion → ①·⁵ artifact cleaning (separate the subject's prose from what they quoted or pasted) → ② local multilingual embeddings [12, 13] → ③ topic discovery with BERTopic and LLM-assisted reduction [14] → ④ multi-facet classification (topic · bond · emotion · period) → ⑤ temporal knowledge graph → ⑥ OCEAN psychometric profile [7, 8] → ⑦ voice via a local LoRA adapter [15] → ⑧ map-reduce synthesis → ⑨ fidelity evaluation. Each stage has a defined interface and is tested in isolation.

Two design decisions govern quality. First: LLM-assisted topic reduction over BERTopic avoids the fragmentation into hundreds of micro-topics that density clustering produces over short, noisy messages —a failure observed in the N1 run—. Second: turn fusion (collapsing consecutive messages from the same sender into a dialogue unit) and artifact cleaning are prerequisites to avoid contaminating the style model with text the person did not write but pasted.

Vector retrieval returns anecdotes by similarity but lacks relational and temporal structure: it mixes chronological milestones and confuses contexts of different people. The fact layer is therefore a directed graph —nodes: people, organizations, places, projects, events, periods; edges: explicit dated relationships— enabling multi-hop reasoning over time. The runtime combines both in hybrid retrieval: the vector supplies the fact or anecdote, the graph supplies the relationship and its evolution. It is the difference between reciting a 2020 message and answering “how did my relationship with a person change between 2019 and 2024,” which the flat store cannot resolve.

This separation is the functional analog of the complementary systems [3]: fast episodic retrieval over the index and graph, plus a background consolidation process —inherited from the generative-agents reflection pattern [16]— that integrates recent conversations into new facts and relationships without saturating working memory.

Attaching a person's style examples to a large model's prompt yields shallow imitation and, often, verbatim regurgitation of the corpus. Training a low-rank adapter (LoRA) on the cleaned personal prose burns the idiolect into the weights: fillers, rhythm, openings and closings, register shifts by interlocutor. Style stops being a retrieved example and becomes a disposition of the model —the move from episodic to semantic described above— and constitutes the first brick of a truly own model, independent of any provider. Traits are captured as a continuous OCEAN vector derived from linguistic markers [7, 8], rather than categorical typologies with no stable psychometric validity.

The methodological differentiator against existing products —all black boxes— is that fidelity is measured, not asserted. Stage ⑨ defines four families of metrics, all local and open. Style: authorship verification by character n-grams with cosine similarity (the PAN task procedure) [11] plus cross perplexity of the adapter over held-out real messages. Consistency: a C-score by natural-language inference (entailment minus contradiction against the biographical profile) and a P-score by distance to the corpus's personality sentences, in the manner of PersoBench. Behavior: decisions in simulated scenarios graded by a calibrated evaluator, in the manner of PersonaGym. Values: prediction of the twin's value profile from free conversation and its correlation with the real subject's, measured with the Schwartz PVQ-RR questionnaire [10].

None of this is asserted as an achieved result: stage ⑨ is specified and pending. The methodological commitment —and the reason the correct name is “measured fidelity” rather than “high fidelity”— is to publish the number, with its interval, once the suite runs over the full corpus.

The engine is not theory: the lightweight version ran end-to-end on real data. The parser normalized 35,023 correspondence messages from 2019 to 2026 without manual tuning, handling the local format's traps (the narrow no-break space U+202F, multiline, attachments, system messages), of which 12,049 are the subject's.

Topic discovery over 4,749 substantive subject messages (k = 12) surfaced coherent axes without imposing categories —coordinating/meeting, providing, the everyday, care and affection, conflict, reflection, family—, ranking dimensions a sampled reading underestimated. Voice synthesis v0 captured the dialect (voseo with a shift to “tú” in affective register), canonical openings and closings, and register shift by interlocutor. Due to a work-machine constraint (Application Control blocks native numpy/sklearn), this run used a pure-Python implementation (TF-IDF + spherical k-means); the heavy stages (embeddings, BERTopic, LoRA) run in a dedicated Linux environment.

The raw corpus —years of intimate correspondence— never enters version control or the cloud during analysis; embeddings, clustering and adapter training are local. The policy, however, is not singular, and that distinction is by design. In personal mode (the owner's own twin) privacy means sovereignty: the history is kept whole —the data is the person; deleting it is loss— but encrypted at rest, with the technical guarantee that the raw never leaks to third parties.

In product mode (a service for third parties) privacy means minimization: the household exemption disappears, the company becomes the data controller, and the messages of contacts who never consented are purged and anonymized at the source, before any analysis —a technical, not contractual, answer—, with explicit consent, the right to erasure, and an impact assessment (DPIA). The same engine serves both policies; what changes is what is kept and where the anonymization boundary sits.

The claims in this document must be read with their scope. The evaluation suite (⑨) has not yet run: fidelity is designed to be measured, not measured yet. Extraction may bias the portrait toward the lens of the densest sources; this is offset by a first-hand curated core and will be quantified by ⑨. A twin of style and biography reproduces how a person expresses themselves and what they remember, not their conscious experience: behavioral fidelity is not personal identity, and the neuroscientific frame is a design principle, not a claim of equivalence with the brain. Finally, the preliminary results come from a subset of the corpus and a lightweight implementation; the figures will be revised when the heavy stages run over the full history.