Context graph freshness is the measured, governed currency of every fact in an enterprise context graph: how closely the Enterprise Digital Twin tracks its source systems, expressed per fact and per tier rather than per batch load. It is maintained through continuous synchronization, change data capture, event streaming, scheduled sync, and re-resolution, and enforced by making fact age visible to the Decision Layer and binding in the Context Harness.
Context graph freshness: why a stale graph is worse than no graph #
Context graph freshness is the difference between a twin and a portrait. The week a context graph goes live, it is the best representation of the enterprise anyone has seen. Six months later, if nothing keeps it synchronized, it is something more dangerous: an authoritative-looking model of a company that no longer exists, feeding AI systems whose confidence never flags. A missing fact makes a model hedge; a stale fact makes it certain and wrong. That asymmetry is why freshness is not an operations detail but a first-order design property, and why CTOs should treat it the way they treat availability.
The failure is gradual, which is what makes it common. Nothing breaks. Sync jobs that ran nightly keep running nightly while the business starts changing hourly. An acquisition adds entities nobody re-resolves. A pricing system migration quietly severs a connector. Each decision the graph supports gets slightly worse, and because the degradation is smooth, no incident ever triggers a review. The remedy is to engineer freshness deliberately: a synchronization portfolio matched to source characteristics, per-fact timestamps, and tiered SLAs with enforcement. Figure 1 shows the pipeline.
The synchronization portfolio #
No single sync mechanism fits an enterprise, so freshness engineering is portfolio management. Change data capture suits transactional systems: row-level deltas flow from database logs with near-real-time lag and no load on the source application. Event streaming suits operational state: telemetry, shipments, and status changes arrive as they happen and update the twin's state component directly. Scheduled synchronization remains right for slow movers: product hierarchies, org structures, and policy documents that change weekly do not justify streaming infrastructure. The design skill is matching mechanism to source volatility, not maximizing real-time coverage for its own sake.
Two subtleties separate a living graph from a merely updated one. First, re-resolution: when an identity-bearing attribute changes, a legal name, a registration number, a merger, the Context Graph Engine must re-run entity resolution for the affected neighborhood, or the graph accumulates identity drift that no delta feed fixes; the mechanics are covered in Entity Resolution Across Enterprise Systems. Second, relationship recomputation: a fact change can invalidate derived edges, a supplier's ownership change alters risk relationships downstream, so the engine propagates consequences rather than just storing the delta. Both are what distinguish maintaining a twin from mirroring tables, a distinction rooted in the difference between records and meaning explored in Context vs Data.
Freshness SLA tiers #
| Tier | Illustrative target lag | Typical facts | Sync pattern | On breach |
|---|---|---|---|---|
| Tier 0: Live | Seconds to a minute | Positions, machine status, availability, fraud signals | Event streaming | Harness blocks dependent decisions; page on-call |
| Tier 1: Operational | Minutes to an hour | Orders, tickets, inventory, account balances | Change data capture | Decisions proceed with staleness flagged in evidence |
| Tier 2: Relational | Hours to a day | Ownership changes, contract amendments, org moves | CDC plus re-resolution triggers | Affected neighborhood marked pending re-resolution |
| Tier 3: Reference | Days to a week | Product hierarchies, policy documents, master data | Scheduled sync | Warning on dashboard; refresh prioritized |
The lags are illustrative ranges, and the point is the structure, not the numbers: tiers are set by decision impact, each fact belongs to exactly one, and the breach column is what turns freshness from a metric into governance. A tier without a consequence is a wish.
Freshness in three industries #
Logistics. A freight operator's twin holds vehicle positions and driver hours at Tier 0 via event streams, shipment orders at Tier 1 via CDC, and carrier contracts at Tier 2. When a rerouting agent proposes a plan, the Harness verifies that position facts are seconds old; a streaming outage does not produce wrong routes, it produces blocked automated rerouting and an escalation, which is the correct failure.
Banking. Exposure and limit facts sit at Tier 1, group ownership at Tier 2 with re-resolution triggers on registry changes. The dangerous case is the quiet one: a subsidiary sale that never triggers re-resolution leaves group exposure overstated or understated for months. Tier 2 breach alerts exist precisely for the changes nobody emails about.
Retail. Store inventory streams at Tier 0 during trading hours, supplier lead times update at Tier 1, and the product hierarchy syncs nightly at Tier 3. The replenishment agent's recommendations carry each input's age in their evidence, so a category manager can see at a glance that a surprising suggestion rests on an hour-old lead-time change rather than a stale assumption.
A realistic enterprise scenario #
A CTO at a global components distributor investigates why the quoting agent, accurate at launch, has drawn rising complaints. The audit finds no defect: every sync job runs green. But the jobs were designed two years ago, nightly for everything, while the business moved to intraday price updates from volatile commodity inputs, and an ERP migration rerouted the pricing table the connector never learned about. The graph is fed, and it is fed yesterday's company.
Understand. The team classifies twin facts into tiers by decision impact: cost and availability to Tier 1 CDC, customer agreements to Tier 2 with re-resolution triggers, catalog structure to Tier 3. Per-fact freshness stamps are added so age is visible, not assumed.
Decide. The Decision Layer now includes fact age in every quote's evidence, and the Context Harness enforces the tiers: quotes on cost facts older than the Tier 1 window are flagged for review instead of sent.
Execute. The Execution Grid issues quotes and writes them back as fresh state. As an illustrative range, teams that move from batch-era sync to tiered freshness typically see stale-input incidents fall to a small fraction of their prior rate within a quarter, and, just as valuably, the remaining staleness becomes visible and attributable rather than silent.
Common mistakes to avoid #
- One freshness number for the whole graph: "synced nightly" hides the fact that some decisions need seconds and others need weeks; tiers exist because facts differ.
- Green pipelines mistaken for fresh context: jobs succeeding is not the twin tracking the business; measure fact age at consumption, not job status at load.
- Streaming everything: real-time sync for reference data burns budget and attention that Tier 0 facts actually need; match mechanism to volatility.
- Skipping re-resolution: applying deltas without re-running identity for changed entities lets the graph drift into duplicate and merged identities no feed will fix.
- Freshness without enforcement: if the Harness cannot block or flag decisions on expired facts, staleness remains invisible exactly when it matters.
- Forgetting the write-back: outcomes the Execution Grid does not record become the freshest facts the twin is missing.
The CTO's freshness checklist #
Freshness is governable in four moves. Stamp every fact, because unmeasured age cannot be managed. Tier every fact by decision impact, with breach consequences defined. Match sync mechanisms to source volatility, CDC and streams where impact demands, schedules where it does not, with re-resolution and relationship recomputation wired in. And surface freshness where decisions happen: in the Decision Layer's evidence, in the Harness's rules, and on the same executive dashboard as availability. A context graph earns the word "twin" only while these four hold; the moment they lapse, it is aging into a portrait.
Freshness is one of four measurable quality dimensions; the full scorecard is the subject of Context Quality: Measuring What AI Knows. For how state fits alongside the other four context components, see The Anatomy of Enterprise Context, and for the engine performing the synchronization, The Context Graph Engine, Explained in the Fundamentals cluster.
How OpenKnowra approaches this #
The tiering and portfolio approach above works on any stack. OpenKnowra builds it in: the Context Graph Engine ingests change data capture, event streams, and scheduled syncs, stamps every fact in the Enterprise Digital Twin with per-fact freshness, triggers re-resolution when identity attributes change, and recomputes affected relationships. The Context Harness enforces freshness tiers at decision time, the Decision Layer includes fact age in every evidence trail, and the Execution Grid closes the loop by writing outcomes back as current state.
A quick diagnostic we offer: point us at one decision domain and we will produce its fact-age distribution, how old the inputs to your AI decisions actually are, which is usually the most persuasive freshness artifact a CTO can bring to an architecture review.