Data models

Context is the flagship model; the rest make the store broadly useful.

TemporalStore is positioned around Context for AI agents: memory streams, tool traces, retrieval evidence, summaries, and replay. Counters, aggregates, distinct sets, sequences, and profiles are supporting primitives for the same online state engine.

TemporalCounter

Bucketed count, sum, rate, quota, and velocity state over sliding or tumbling windows.

Use cases

Fraud velocity checks, API rate limits, ad frequency caps, abuse throttles, cart activity, and agent safety counters.

TemporalAggregate

Filtered min, max, sum, count, grouped rollups, and numeric feature windows by entity.

Use cases

Risk features, campaign spend windows, recommendation freshness, operational health rollups, and marketplace scoring.

TemporalDistinct

Approximate or exact distinct entities observed inside a time range.

Use cases

Unique devices per user, merchants per card, IPs per account, sessions per workspace, and unique tools used by an agent.

Sequence

Ordered behavior history with event-time queries, tail reads, and replay-friendly evidence.

Use cases

Product event timelines, investigation trails, ranking sequences, shopping journeys, and agent tool-call history.

Profile

Latest entity facts colocated with temporal state so online decisions have stable identity context.

Use cases

User profile state, account risk level, workspace configuration, device metadata, tenant policy, and model flags.

Context

Structured agent memory streams, retrieval traces, summaries, decisions, and supporting evidence.

Use cases

Open-ended agent memory, iterative coding sessions, human approval history, tool traces, ContextPacks, safety counters, and reproducible AI decisions.

How to read the model set

Each model should have tests, benchmarks, and operational evidence.

Benchmarks Measure model behavior under realistic write/read mixes. Latency, throughput, correctness, and context quality. Repository Track implementation and source artifacts. Rust engine, harnesses, docs, and reports. API Expose compact primitives without losing semantics. Ingest events, read windows, and build context.