Sophia’s design centers on development through experience: retaining knowledge, revising expectations, and acquiring a disposition that influences later perception. LOGOS divides the supporting work among Sophia as the cognitive core, Talos for sensing and action, Hermes for language services, and Apollo for interaction and inspection.
The earlier implementation uses a graph to connect entities, relations, types, and records about Sophia herself. A persona diary gives beliefs, decisions, observations, and reflections a persistent history. The larger design brings abstract knowledge, grounded experience, and affective state into a common representation while allowing them to change at different rates.
Cliste is intended to become Sophia. Its vector-symbolic work explores how content can be bound to context and recombined within the representation itself. Existing Sophia interfaces can change around that work; compatibility with the earlier graph system is not a constraint on the research.
Why this approach
Memory across different timescales
The memory design distinguishes ephemeral state, short-term experience, and long-term knowledge. Ephemeral state includes the active situation and passing mood. Short-term records keep an encounter available for reassessment. Long-term memory retains knowledge and dispositions that should continue to matter after the immediate situation has passed.
Salience would help determine what survives, with later reflection able to reconsider an earlier judgment. Repeated experience could contribute to a lasting disposition while a temporary reaction fades. This is the intended retention model: the current code provides in-process state and persistent records, but does not yet implement the complete three-tier promotion, expiration, and forgetting cycle.
Internal state changes perception
My affect hypothesis is that internal state changes how an input is perceived before the system reasons about it. Recent experience and a longer-lived disposition could alter what stands out, which associations become accessible, and how a new encounter is interpreted.
Separate Sophia experiments implement this idea as a matrix transform over embeddings, including update rules that let new inputs change the matrix. Connecting it to the persona diary, temporary mood, and useful changes in continuing behavior remains work to do.
Structure has to help the individual use experience
A richer graph is useful if it helps Sophia recover relevant experience, explain a difference, or anticipate a consequence. The graph studies showed that stronger similarity grouping can reflect a shared source passage, and that repeated maintenance can create more types without simplifying the account.
Those findings motivated the move toward Cliste and its investigation of a common substrate for perception, knowledge, affect, and relationships. Actions provide consequences against which the resulting expectations can be tested and revised.
Where it stands
Sophia's persona API stores beliefs, decisions, observations, and reflections with timestamps, confidence, sentiment, and optional links to goals or processes. It supports filtered retrieval, history-preserving updates, and summaries of recent sentiment. Apollo's browser diary reads this API, and its chat brings up to five recent entries into the language model's context.
The persistence layer uses a shared state envelope with provenance for abstract, grounded, and emotional records, identified as CWM-A, CWM-G, and CWM-E. These distinguish kinds of content; the ephemeral, short-term, and long-term design distinguishes how long content remains useful. Shared storage does not by itself implement transfer or learning among those systems.
Graph ingestion, type emergence and rollup, placement checks, and scoped inspection are implemented. Sophia exposes graph counts, types, neighborhoods, and search so clients can inspect the represented world through the service.
The development-branch embodiment interface reads observations and sends velocity commands through Talos. Its present actions follow a fixed schedule. Cliste's continuing vector-symbolic memory is the active route toward a core that can acquire and use experience more flexibly.
An entity-representation study compared six ways of encoding 5,330 entities. With names alone, 12.1% of the measured nearest neighbors came from the same source passage. Sentence-based variants raised that to 57.1–60.2%, against a random expectation of about 0.3%.
Adding context produced strong passage grouping. That can be useful for recovering a shared source, but it did not validate a taxonomy of the entities. The study measured passage grouping, not category quality.
Type creation and consolidation are different capabilities
A category-formation test produced 21 types with noun-headed names. A second maintenance pass over the same data retained all 21 identities and created 30 more, taking the total to 51 without ingesting new documents.
The system could form and place categories, but repeated work did not converge toward a compact account. The current emergence code reuses existing names and the rollup code groups types under broader types; neither result establishes that redundant meanings are consistently merged. Consolidating redundant categories remains unresolved.
The embodiment interface exists; adaptive control remains separate
The development-branch embodiment interface attaches Sophia to a body, reads observations, and emits motion commands. Its command changes after ten steps regardless of the observation. Actions conditioned on perception and learned consequences remain to be implemented.
Selected recorded studies. Each keeps its original conditions, date, interpretation, and limits.
Historical experiment · 16 June 2026
When context groups entities by their source passage
This study compared ways of representing 5,330 extracted entities with text-embedding-3-large, using 3,072 dimensions. The question was whether adding context produced more useful entity neighborhoods.
The diagnostic measures the fraction of the ten nearest neighbors that share an entity’s source passage. It samples 3,000 query entities with a fixed seed. The enrichment column compares that fraction with the unrounded random-neighbor expectation.
Recorded source-passage clustering across all six arms
Representation
Neighbors from same passage
Enrichment over random
Neighbor margin
Entity name
12.1%
37.4×
1.107
Containing sentence
57.1%
177.0×
1.000
Name + passage
58.3%
180.8×
1.252
Entity marked in passage
60.2%
186.6×
1.276
Generated gloss
15.2%
47.3×
1.128
Name + gloss
14.8%
46.0×
1.114
Passage context dominated the grouping
Context can make neighbors look more coherent because they came from the same passage. The passage-based arms show this strongly. That makes the apparent improvement a poor basis for claiming better types.
Limits
This is a historical diagnostic of a particular representation pipeline. The intrinsic metrics were not validated as a taxonomy score. It does not establish that embeddings lack useful structure or that graph reasoning is generally superior.