Give agents
a world to
navigate.

A world is a collection of events, their history, and a model that learns how those histories evolve. You and your agents share that context, so each question can start with what has already happened.

A database organizes records for retrieval and computation. Spacefolder adds a learned representation of change over time and trained readouts for outcomes. The architectural goal is to reuse that representation for context and scoring, reducing repeated feature-building and retrieval work in the agent’s prediction path.

Yes. PostgreSQL supports user-defined functions and extensions; ClickHouse includes machine-learning functions. The difference here is the organizing principle: a shared temporal representation and its trained readouts, rather than a general-purpose query engine. PostgreSQL extensibility · ClickHouse model functions

Vector search usually ranks stored embeddings by similarity. Spacefolder’s sequence encoder represents ordered histories, and its prediction readouts learn an outcome from those histories. Similarity is useful for navigation; an outcome score comes from a trained readout, not from nearest neighbors alone.

The recorded run shows roughly 11.9k event operations per second while churn scores update for 150 histories. That figure measures ingestion throughput; prediction latency and accuracy are separate metrics. Watch the run

Keep the tools your workload needs. Latent states are learned summaries, not lossless records. Source events still support exact filters, audit trails, and evaluation; transactional SQL and constraints remain database work. Spacefolder centers the prediction path on the model, while exact retrieval can use conventional storage and indexes.

A readout needs a defined task and suitable training. Reusing an encoder can make it useful across several routes, but a new outcome or domain may require labels, a new readout, or retraining. Instruction conditioning selects or shapes supported scoring behavior; a prompt alone does not create a validated predictor.

The supplied engine design uses a two-layer, 192-dimensional sequence encoder. Last-state, masked-mean, and masked-max pooling form a 576-dimensional summary. A projected instruction embedding can condition that summary before a trained readout. A separate training objective predicts the next latent state from state and action, with SIGReg regularization.

New events extend a history and trigger encoding and scoring work. Keep history, model, and readout versions together so a result can be traced back to its inputs. A new encoder may require re-encoding; embeddings from different versions are not automatically compatible. Evaluate each route on held-out data and track calibration and drift against later outcomes.

The architecture separates an active runtime from durable storage. A Mac, edge host, or cloud host performs ingestion, encoding, and scoring. S3-compatible storage backs versioned histories and model snapshots. The bucket stores data; the runtime performs inference. The recording shows the local scoring loop; the diagram describes the wider runtime and storage design. Memory, hardware, and model size determine the useful working set and cost.

A prediction can help prioritize attention. Establishing an action’s causal effect requires appropriate data, experimental design, and evaluation. Instruction-conditioned scores and a next-latent training objective do not, by themselves, establish a reliable intervention simulator.

Yes. The name is a nod to Dune, where Navigators use foresight to plot safe paths while a ship’s engines fold space. We borrow that relationship between prediction and navigation: Spacefolder folds event histories into a learned model space, and agents act as Navigators, querying its trajectories and using predictions to plan their next move. The model folds the data. The agents chart the course.