What Is Live Context?
Live context is decision context kept close to the current state of the world. Reads run against the latest committed data, and derived state — aggregates, counters, features, views — is refreshed incrementally, each refresh applying only what changed, on a schedule typically every few seconds, rather than recomputed from scratch by a batch job. A decision reading live context reads the world approximately as it is now, not as it was at the last batch run or cache refresh.
Updated
What is live context?
Live context is one of the three capabilities that define a Context Lake in Tacnode's vocabulary, alongside shared context and semantic context. It addresses the freshness half of the context gap: the pipeline lag between an event occurring and the derived state that decisions read reflecting it.
Two mechanisms carry it. On-demand reads compute a signal at query time against the latest committed data, so there is no pipeline between the event and the read; for most signals this is the whole answer. Incremental maintenance covers the aggregates too expensive to compute per request: instead of rebuilding a view from scratch — the model of batch jobs and standard materialized-view refresh, whose cost grows with data volume — each refresh applies only the changes since the last one. Because a refresh costs what changed rather than what exists, it can run every few seconds. Incrementally maintained views are explained in incremental materialized views; the broader argument is the live context pillar and live context: the key that unlocks real-time AI.
Freshness is set per signal. A view's freshness is its refresh schedule, written in SQL; a signal that needs to be current to the last committed write is read on demand.
How live context works
Compare the three ways to keep one velocity counter fresh:
- Batch recompute — a job rebuilds the counter hourly; staleness up to the full interval, cost proportional to all data scanned.
- Incremental refresh — the counter's view refreshes on a short schedule, typically every few seconds; each refresh applies only the new events, so cost tracks the delta and the schedule can be tight.
- On-demand read — the decision computes the counter at query time against the latest committed data; no refresh interval at all.
The combination is what makes narrow validity windows meetable for derived signals: a 200-millisecond authorization reads the signals that must be current on demand, and reads the heavy aggregates from views refreshed every few seconds.
Why live context matters
Decisions inherit the age of the freshest context they can actually read. Models, rules, and agents evaluated against state minutes or hours old are structurally blind to exactly the events that most need a response — the burst, the spike, the concurrent draw-down — because full-recompute pipelines lag most when event volume is highest. Live context shrinks that distance: on-demand reads see the latest commit, and incremental refresh keeps aggregates seconds behind instead of a batch interval behind. For AI agents, whose multi-step plans compound on each read, the difference between batch-fresh and live context is the difference between reasoning about the world and reasoning about a recording of it.
FAQ
Related terms
A Context Lake is real-time, multi-modal context infrastructure that gives AI agents and decision systems fresh, internally coherent context at decision time.
A context gap is the difference between the state of the world an automated decision acts on and the actual state at the moment the decision commits.
Derived context is state computed from events rather than stored directly — aggregates, velocity counts, features, embeddings — that decisions consume.
A validity window is the interval within which a decision's context remains an accurate basis for action — typically 10ms to 1s for automated decisions.
Shared context means every service and agent in a decision system reads one internally coherent snapshot of state, instead of private copies that diverge.
Semantic context is the meaning-bearing layer of decision context — aggregations, vector similarity, LLM-derived signals — computed from the same events.