A single memory unit stored in the memory system.

Configuration for the memory system.

An agent with persistent semantic memory. Every call to remember() persists to disk. Every call to recall() searches semantically across

Generates embeddings for text.

Persistent semantic memory store. Thread-safe SQLite storage with embedded vector search. All operations are atomic — no corrupted state

Structured operational memory for AI agents. Built on top of Agent memory but with purpose-built methods for recording and querying exec

Automatic learning collector. Ingests execution events from ARK Runtime and records them as structured experience. Supports three ingest

ObservationKind describes what was observed.

Result of a recall query — a memory plus its relevance score.

CodeBlock represents an extracted code block from a response.

CodeQuality holds quality metrics for a code response.

DisagreementResult holds the outcome of comparing two model outputs.

EffortPlan describes the compute budget for a task.

═══════════════════════════════════════════════════════════ Event Emitter Writes structured JSONL events during execution. ARK Memory's Collector ing

FailurePrediction is the Teacher's instinct about whether a model will fail.

FuncParam holds a function parameter.

FuncSignature holds a parsed function signature.

Governor holds the verifier and registry — the "Teacher" layer.

GovernorRegistry is the interface the agent uses to record model observations.

GovernorVerdict is the simplified verdict the agent sees.

GovernorVerifier is the interface the agent uses to verify outputs.

Observation is a single data point the registry records.

Registry is the Model Capability Registry — the Teacher's knowledge of every student.

StepTypeRecord tracks how a model performs for a specific step type.

ToolRecord tracks how a specific model performs with a specific tool.