Flagship data model
Context Management: replayable short-term and long-term memory for AI agents.
Context is the reason TemporalStore exists. It keeps everything agents, teams, and devices need to remember as time-indexed records — so the model gets the right short-term and long-term context at the right moment, and you can replay exactly why any decision was made.
What agent context is
Cross-session, cross-agent, cross-device memory in one temporal record type.
An agent's memory is not one blob of text. It is a stream of distinct, timestamped facts that each age differently across sessions, agents, devices, users, and teams. Context Management stores all of them as the same time-indexed primitive, so short-term working memory and long-term team memory can be filtered, summarized, and replayed together.
- Session and device memory — user goals, messages, edits, files, device context, and long-running task state.
- Cross-agent team memory — shared decisions, project facts, handoffs, and preferences that survive across agents and sessions.
- Tool traces — commands, tool calls, approvals, generated artifacts, and failures, so an agent can resume with evidence instead of guessing.
- Retrieval evidence — source-backed memories with timestamps and relevance signals, packed into a compact context pack.
- Summaries — rolled-up L0/L1 digests that keep long histories small while staying source-linked.
- Safety counters & replayable decisions — rate limits, policy events, and escalation history beside the memory stream, so you can reconstruct the exact context an agent had when it answered.
How it works
Events flow in; a fresh, bounded context pack flows out.
Every message, tool call, and decision is appended as an event — a cheap, ordered write. At request time the engine assembles a context pack: it pulls the relevant streams, drops superseded and duplicate records, folds old history into summaries, and trims to a token budget.
Because assembly happens in the store rather than in your prompt buffer, the pack is fresh (reflects the latest event), de-duplicated, source-backed (each memory carries its origin), and replayable (ask for the pack as-of any past moment). The result is fewer tokens spent on stale or repeated context, and better answers grounded in evidence.
The store does the assembly, so the agent receives evidence instead of an ever-growing transcript.
In practice
Build context, write a memory, and block stale reads.
context = ts.context(
entity="workspace_7",
streams=["tool_calls", "decisions"],
since="24h",
summarize=True,
include_evidence=True,
)ts.put_event(
table="agent_memory",
entity="workspace_7",
ts_ms=now_ms,
attrs={"kind": "decision", "text": "...", "source_ref": "pr#42"},
)pack = ts.context(
entity="workspace_7",
as_of="request_time",
drop_superseded=True,
token_budget=4096,
)When to use it
Reach for Context Management when memory has to be defensible.
Use it whenever an agent must remember across turns or sessions, resume a long task, or justify a decision after the fact. It beats stuffing history into the prompt (which grows unboundedly and repeats itself), a plain vector index (no time order, no supersession, no audit trail), and a cache (no replay, no evidence links).
Context composes the supporting primitives below: ordered history from Long Sequence Feature, safety counters and distinct sets from Control State, and rollups from Aggregated Feature.