Data models
Context is the flagship; the rest make the store broadly useful.
Each data model is a time-indexed primitive the engine understands directly — not a table you assemble by hand. Open a detail page for depth, or build your own when a workload needs one.
The data models
One flagship, composed from a handful of temporal primitives.
Every model stores per-entity state indexed by time, so the same reads — tails, windows, filters, and replay — work across all of them. Context Management is the model most teams reach for; underneath, it composes the supporting primitives, and each of those is useful on its own. All of it is persistent and memory-first: hot state serves from RAM while the engine moves data automatically across memory, SSD, and shared storage — durable by default, without a separate cache tier to keep in sync.
Each node is a first-class object in the engine; Context stitches them into agent memory.
Model detail pages
Open the three supporting data models directly.
These are not side notes. Long Sequence Feature, Aggregated Feature, and Control State are the reusable temporal primitives that make Context Management work, and each has a dedicated page with workload shape, read/write behavior, and examples.
Serve long, ordered per-entity histories for ranking, personalization, agent traces, click streams, and recent behavior recall.
Read the page Windowed rollups Aggregated FeatureMaintain high-cardinality sums, counts, min/max, grouped filters, and fresh risk or campaign features over moving windows.
Read the page Limits and state Control StateTrack frequency caps, velocity checks, counters, distinct fan-out, cooldowns, and policy state directly in the serving path.
Read the pageAt a glance
What each model stores, and when to reach for it.
| Model | What it stores | Example use |
|---|---|---|
| Context Management | Session memory, tool traces, evidence, summaries, safety counters | Replayable memory for an AI agent |
| Long Sequence Feature | Long, ordered per-entity event history | Last N interactions for a ranker or agent |
| Aggregated Feature | Filtered sum/min/max/count & grouped rollups over windows | Fresh risk or campaign-spend features |
| Control State | Bucketed counters, frequency caps, velocity checks | Rate limits and ad frequency caps |
| Distinct Set | Unique devices, merchants, IPs, sessions in a window | Fraud fan-out and abuse detection |
| Profile | Latest entity facts colocated with temporal history | Stable identity context beside the timeline |
Extensibility
Build your own data model for your workload.
TemporalStore data models are engine-aware objects with explicit commands over entity-local state. Adding a new data model is a small, testable recipe — not a fork of the engine. Enterprise backends like MatrixObject Enterprise and MatrixKV Enterprise plug in underneath without changing the model code.
Define the object
Describe the per-entity state and its encoding — what lives in memory, dumps into pages, and recovers from WAL.
Declare commands
Add the model-aware write and read commands — append, increment, window, distinct — the engine runs at the shard.
Wire indexes & dumps
Register bucket and page bookkeeping plus any secondary indexes so dirty state persists and reloads correctly.
Add tests & benchmarks
Cover correctness, reload and replay, and a workload benchmark — then it is a first-class model in the store.