Give agents
a world to
navigate.

A world connects observations, actions, and outcomes over time. It includes evidence from external systems, the choices your agents make, and what happens afterward. Agents can search that history and use trained predictors to plan their next move.

A world model learns how events unfold over time. Spacefolder uses that learned history to score possible next steps and give parallel agents shared context. Predictions, choices, and measured outcomes stay connected to their source evidence.

Yes. Postgres supports custom functions, and ClickHouse includes model training and evaluation functions. Spacefolder brings the sequence model, trained predictors, and agent decisions into one shared world. Exact source records remain available alongside the learned state.

Vector search retrieves similar items. Systems such as Turbopuffer also combine it with full-text search and filters. Spacefolder learns from ordered histories and uses trained predictors to score outcomes. Agents use those predictions to compare next moves against relevant evidence.

Agent notes, tool calls, and patches belong in the history. So do CI failures, sensor readings, inspection results, and conversions from systems outside the agents. Spacefolder is designed to connect decisions to that external evidence and to the outcomes you can measure.

The design gives parallel agents a shared, versioned snapshot to work from. Each records its proposed action, supporting evidence, predictor version, and expected outcome. Other agents can inspect the reasoning before dividing the work. They compare results against the same evidence. Scheduling, permissions, and conflicting actions need explicit coordination as well.

Claude Code, Codex, Cursor, Grok Bot, internal research agents, robot agents, and custom agents. Each of these connections needs an adapter, authorized access, and verification against the supported tools and questions.

The recording shows ingestion throughput above 11,000 event operations per second. During the same run, the dashboard updates churn scores in batches of 150. Throughput, inference time, and prediction accuracy measure different things. Watch the run

Each predictor needs a supported question, suitable examples, and a measurable outcome. Small readouts can share an encoder, but a new task may need labels and further training. An instruction selects or conditions a supported readout. Training still needs examples and observed outcomes.

A trained predictor can score an action when its training data supports that question. The model foundation also includes a next-state dynamics head. Multi-step simulation remains a development direction. A predicted trajectory doesn’t establish that an intervention causes an improvement. That needs controlled experiments and observed outcomes.

The documented foundation 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 scores an outcome. During training, a separate dynamics head predicts the next latent state from the current state and action. SIGReg regularizes the representation.

The recorded engine runs locally. The product design starts on Mac, with edge and cloud runtimes and S3-compatible shared storage planned. Versioned snapshots identify the exact history and models behind a decision. Runtimes must manage permissions and resolve conflicting updates when synchronizing. Object storage holds the snapshots.

Yes. Dune’s Navigators use foresight to plot safe routes while spacecraft fold space. In Spacefolder, folding means encoding event histories into a compact model space. Agents act as Navigators by querying those trajectories and using predictions to plan their next move.