| Dimension | Detail |
|---|---|
| What it is | A dynamic graph linking data assets, decisions, and temporal context |
| How it differs from knowledge graphs | Adds provenance, governance, and decision traces |
| Key technique | Reification — treating relationships as first-class objects |
| AI agent use | Persistent memory, reasoning substrate, audit trail |
| Key standard | W3C Context Graphs Community Group (est. Feb 2026) |
| Gartner projection | 50%+ of AI agent systems will use context graphs by 2028 |
Where do context graphs originate from?
Permalink to “Where do context graphs originate from?”A context graph is a dynamic network linking data, decisions, and time. Unlike knowledge graphs, it preserves the why behind data relationships, not just the what. Context graphs add temporal context, decision traces, and agent memory to data assets. According to Gartner’s Emerging Tech: Context Graphs Shape the AI Vendor Landscape report, over 50% of enterprise AI agent systems will use context graphs by 2028.
The concept was first formalized in a 2024 Cornell University paper on expanding knowledge graphs with contextual information, and gained industry momentum when Foundation Capital framed context graphs as “AI’s trillion-dollar opportunity” in December 2025.
The core innovation is temporal depth. A context graph does not just answer “what is related to what.” It answers “how did this relationship come to exist, when did it change, and why?”
Context graphs are the underlying data structure for context-aware AI agents, active data governance, and audit-ready data products. Knowledge graphs are static snapshots. Ontologies define formal rules. Context graphs record what actually happened and keep recording as conditions change.
Three-way comparison: context graph vs. knowledge graph vs. ontology
Permalink to “Three-way comparison: context graph vs. knowledge graph vs. ontology”| Dimension | Context graph | Knowledge graph | Ontology |
|---|---|---|---|
| Primary purpose | Operational memory and AI grounding | Entity relationship mapping | Formal concept taxonomy |
| Temporal context | Yes, tracks change over time | Limited | No |
| Decision traces | Yes, first-class objects | No | No |
| Governance | Policy-embedded, automated | External policies | Embedded in rules |
| Relationships | Dynamic, evolving | Static or semi-static | Schema-defined |
| AI agent readiness | Native: reasoning and audit | Partial: retrieval (RAG) | Reference layer |
| Key limitation | Requires active metadata infrastructure | No decision memory | Static, no temporal |
For a full architectural comparison, see context graph vs. knowledge graph.
How context graphs differ from knowledge graphs. Image by Atlan.
How do context graphs work?
Permalink to “How do context graphs work?”Context graphs are built from six core components that work together to produce capabilities knowledge graphs and ontologies cannot.
- Nodes — any entity with operational significance: data assets, people, policies, business processes. Each node carries attributes describing the entity at a given point in time.
- Edges — typed, directed relationships between nodes (e.g., “is certified by,” “depends on,” “is governed by”). Edge type selection is deterministic.
- Temporal attributes — every node and edge carries
valid_fromandvalid_totimestamps, enabling point-in-time queries like “What was the certification status of this dataset on January 15?” - Reification — the technique of treating an edge as a node in its own right, giving it attributes: who created it, when, under what authority, and what caused it to change. Reification makes decision traces possible.
- Active metadata — rather than manual curation, active metadata pipes events (query executions, certification changes, ownership transfers, schema modifications) directly into the graph. The graph evolves in real time.
- Decision traces — the resulting audit trail. Every governance action, pipeline run, and classification event becomes a queryable record. A data engineering leader can answer: “Why was this column reclassified as PII three months ago, and who approved that change?”
Data lineage integrates naturally as a subgraph. Column-level lineage edges carry temporal context. A transformation that changed a column’s data type on a specific date — say, when a dbt model converted recognized_revenue_q4 from FLOAT to DECIMAL(18,2) — is queryable against the governance context of the same time window.
What is the difference between context graphs, knowledge graphs, and ontologies?
Permalink to “What is the difference between context graphs, knowledge graphs, and ontologies?”Context graphs, knowledge graphs, and ontologies serve different architectural roles. The terms overlap in vendor marketing, but the structural differences are precise.
Context graph vs. knowledge graph — key differences
Permalink to “Context graph vs. knowledge graph — key differences”A knowledge graph maps entities and their relationships. Google’s Knowledge Graph connects “Paris” to “France” to “Eiffel Tower.” It answers factual questions about static relationships but has no record of when relationships changed, who established them, or whether a governance policy treats them as valid.
A context graph adds three layers knowledge graphs lack: time-aware relationships, a record of who decided what and when, and governance policies embedded directly in the graph structure. In AI applications, this distinction matters at the retrieval layer. GraphRAG uses knowledge graph structure to improve retrieval-augmented generation. Context graphs go further, providing temporal and decision history that lets an agent reason about data trustworthiness, not just data content.
Context graph vs. ontology — when each applies
Permalink to “Context graph vs. ontology — when each applies”An ontology defines a formal concept taxonomy. It specifies that a “Customer” is a subtype of “Person” and that relationships follow specific logical constraints. An ontology is static by design — no memory of past states, no mechanism for recording reclassifications or ownership transfers.
Context graphs use ontological definitions as their schema layer. The ontology defines valid relationship types; the context graph records actual instances with full history of when relationships changed and who authorized those changes. In mature data governance programs, all three layers coexist: ontologies define meaning, knowledge graphs instantiate entities, and context graphs track operational history.
A semantic layer sits adjacent, translating business metrics into queryable definitions. The semantic layer answers “what does revenue mean?”; the context graph answers “who changed the revenue definition, when, and which downstream reports were affected?”
When to use all three together
Permalink to “When to use all three together”Use an ontology for shared vocabulary and logical constraints. Use a knowledge graph for entity discovery, relationship traversal, and retrieval augmentation. Use a context graph for audit trails, governance automation, temporal queries, or AI agents that must explain their reasoning. Most enterprise data programs benefit from all three, layered in that order.
Why do AI agents need context graphs?
Permalink to “Why do AI agents need context graphs?”AI agents fail predictably without context graph infrastructure. They hallucinate relationships that no longer exist, apply outdated governance policies, cannot explain recommendations, and lose decision memory between sessions. Context graphs address all four failure modes structurally.
Hallucination reduction. An AI agent querying a knowledge graph may retrieve a relationship valid six months ago but since revoked. A context graph exposes valid_from and valid_to timestamps, so an agent filters to currently valid data before reasoning.
Stateless reasoning. Agents lacking persistent memory treat each query as independent. A context graph provides the session-spanning memory layer for consistent reasoning across interactions.
Decision memory. When an agent recommends a data asset for model training, that decision must be retrievable on the next run. Context graphs store decision traces as first-class objects, creating the feedback loop that makes agents improvable.
Audit and explainability. Regulators require AI systems to explain recommendations. A context graph enables post-hoc audit: “This agent recommended Dataset X because it was certified on March 10, had no open quality incidents, and had been used in three prior approved training runs.”
The context layer for AI describes how context graphs slot into AI agent infrastructure. The Model Context Protocol (MCP) is the transport mechanism. It defines how agents request and receive context. Context graphs are the data structure those systems expose via MCP.
How do context graphs reduce token costs for AI agents?
Permalink to “How do context graphs reduce token costs for AI agents?”Large language models charge by the token. Passing full documents into every agent prompt burns through token budgets fast, especially when most of the document is irrelevant to the query at hand.
Context graphs solve this by delivering pre-structured subgraphs instead of raw text. The agent receives only the relevant entities, their relationships, and applicable policies — no filler paragraphs, no duplicate definitions, no tangential metadata. When context arrives as typed relationships rather than unstructured text, GraphRAG techniques produce more accurate grounding with significantly fewer tokens consumed.
Why does context compound as an AI advantage?
Permalink to “Why does context compound as an AI advantage?”Models are commoditizing fast. GPT-4 held its lead for about six months before Claude, Gemini, and open-source alternatives closed the gap. Model intelligence is table stakes.
Context is different. Your enterprise’s accumulated understanding of what data means, how decisions get made, and what “revenue” actually means in your specific business does not commoditize. It compounds. Each agent interaction that terminates in a decision — “Dataset X was certified for training use on March 12” — becomes a queryable trace. The fiftieth query against a mature context graph benefits from forty-nine prior decision traces that the first query had no access to. The model is the same. The context is richer.
For teams building an AI-ready data catalog, context graphs are the layer that converts a passive data catalog into an active memory system.
GraphRAG vs. context graphs
Permalink to “GraphRAG vs. context graphs”The distinction matters. GraphRAG is a retrieval technique that uses graph structure to improve context passed to a language model. Context graphs are the persistent operational memory that GraphRAG queries. GraphRAG without a context graph retrieves facts. With a context graph, it retrieves facts plus governance history, recency, and trustworthiness signals.
Who is building and investing in context graphs in 2026?
Permalink to “Who is building and investing in context graphs in 2026?”Interloom raised $16.5M in seed funding, Foundation Capital published a “trillion-dollar opportunity” thesis, and the W3C formed a dedicated community group — all within a four-month window in late 2025 and early 2026.
Investment signals
Permalink to “Investment signals”Interloom raised $16.5M in seed funding on March 19, 2026. DN Capital led the round, with Bek Ventures and Air Street Capital also participating. Interloom builds context graph infrastructure for regulated industries, with early customers including Commerzbank, Zurich Insurance, JLL, and Volkswagen. Enterprise adoption from financial services and insurance signals that buyers treat context graphs as production infrastructure, not research projects.
Foundation Capital published its thesis on December 22, 2025, describing context graphs as “AI’s trillion-dollar opportunity.” The thesis frames decision traces as the core primitive making AI agents trustworthy at enterprise scale — the same design principle active metadata platforms like Atlan have built toward since before “context graph” entered wide circulation.
Standards momentum
Permalink to “Standards momentum”The W3C Context Graphs Community Group launched on February 24, 2026. It launched with 56 members. Ron Itelman chairs the group. The charter covers standardizing interchange formats and query semantics. W3C community group formation is the same mechanism through which JSON-LD, RDF, and SPARQL matured into core knowledge graph infrastructure.
Analyst projections
Permalink to “Analyst projections”Gartner’s March 2026 projection that 50%+ of enterprise AI agent systems will incorporate context graphs by 2028 applies specifically to deployments requiring governance, audit, and explainability. Gartner predicts 50% of AI agent deployment failures by 2030 will trace back to insufficient governance platform enforcement. The two-year window gives enterprise architects time to evaluate and implement before the technology becomes a baseline expectation.
Gartner projects 50%+ enterprise AI agents will use context graphs by 2028 for governance and explainability. Image by Atlan.
Practitioner perspective
Permalink to “Practitioner perspective”Dharmesh Shah, co-founder of HubSpot, published a balanced analysis on January 9, 2026. Shah framed context graphs as a solution to agents lacking organizational reasoning context, while expressing skepticism about implementation readiness. Context graphs are architecturally sound and increasingly necessary, but implementations require active metadata infrastructure most organizations do not yet have.
Where do context graphs fall short?
Permalink to “Where do context graphs fall short?”Context graphs are hard to implement. The obstacles are worth being direct about.
“Incremental rather than transformational” — Verdantix (January 13, 2026)
Permalink to ““Incremental rather than transformational” — Verdantix (January 13, 2026)”Verdantix analyst Aleksander Milligan published Context Graphs: Transformational Architecture Or Familiar AI Hype?, characterizing the distinction between knowledge graphs and context graphs as “incremental rather than transformational” and describing context graphs as a “transitional technology.” The assessment frames context graphs as useful intermediate architecture while AI infrastructure matures, but not yet core platform technology. Verdantix raised three concerns: most enterprises lack active metadata infrastructure, query performance at scale is unproven, and absent standards create lock-in risk.
Each concern has a counterpoint. The infrastructure floor has dropped. An organization running Snowflake, Databricks, or BigQuery against an active data catalog already emits the event stream context graphs require. On query performance, context graphs run over metadata, not rows — at metadata scale, current graph databases handle relevant query patterns adequately. On lock-in, the W3C Context Graphs Community Group was formed specifically to address interchange formats.
Implementation complexity
Permalink to “Implementation complexity”A context graph is not a product you buy and turn on. It requires an active metadata layer, a graph storage backend, a query interface, and integration with upstream data platforms — Snowflake, Databricks, BigQuery, dbt, Tableau, or whichever tools your team runs.
The active metadata approach reduces this floor substantially. When a data platform already emits metadata events, those events feed the context graph without additional instrumentation. Organizations with passive, manually-curated metadata face the hardest implementation. Organizations running active metadata platforms face the easiest.
Implementation phases
Permalink to “Implementation phases”| Phase | Focus | Timeline | Key actions | Success metrics |
|---|---|---|---|---|
| 1 | Modern metadata foundation | 2-4 weeks | Deploy active metadata catalog, connect pre-built connectors | Centralized search adoption |
| 2 | Lineage capture and enrichment | 2-3 months | Blend capture methods, build graph-native lineage | Impact analysis time reduction |
| 3 | Semantic layer and governance | 4-6 months | Integrate glossaries, implement policy nodes | Search accuracy, policy compliance |
| 4 | AI activation and agent integration | 6+ months | Enable graph-grounded RAG, connect to BI tools | AI answer accuracy, compliance rates |
Evaluation criteria: when a context graph is the right choice
Permalink to “Evaluation criteria: when a context graph is the right choice”A context graph is the right architecture when:
- AI agents must explain their recommendations with traceable reasoning
- Data governance requires audit trails for regulatory compliance (GDPR, CCPA, EU AI Act)
- Data assets change frequently and consumers need point-in-time trustworthiness signals
- Multiple AI agents must share memory across sessions
A knowledge graph alone is sufficient when:
- The use case is entity discovery and relationship traversal without governance requirements
- Relationships are stable over time and do not require temporal queries
- The primary application is retrieval augmentation, not governance or audit
An ontology alone is sufficient when:
- The requirement is shared vocabulary and logical constraints, without operational tracking
- The system generates or validates data against formal rules
Most production AI governance programs need all three layers. Starting with ontology definitions before layering knowledge graph instances and context graph history is the path with the fewest rework cycles.
How Atlan implements context graphs as a context layer
Permalink to “How Atlan implements context graphs as a context layer”Atlan builds context graphs as a product, not a prototype. The architecture demonstrates what “active metadata” means in operational terms.
Atlan’s data catalog continuously ingests metadata events from connected data platforms: Snowflake, Databricks, BigQuery, dbt, Tableau, and 100+ integrations. Each event (a column certified, a policy applied, an ownership transferred, a query executed, a quality check failed) becomes an edge in the context graph with a timestamp, an attributed actor, and a relationship type.
The result is a context graph that builds itself. Column-level data lineage edges carry temporal context. Governance policy applications are edges. Certification events are edges. Every event that changes operational state is recorded as a graph edge — a complete decision trace of the data environment.
The context layer for AI is Atlan’s product surface exposing context graph queries to AI agent systems. An agent building a model training dataset queries Atlan’s context layer to determine whether candidate datasets are certified, under active governance, and free of open quality incidents.
For organizations evaluating how a context graph fits into their existing data governance program: connect your data platforms, activate active metadata, and the context graph begins building from the event stream. No batch ingestion job, no manual graph construction.
Frequently asked questions
Permalink to “Frequently asked questions”What is a context graph in simple terms?
Permalink to “What is a context graph in simple terms?”A context graph is a network that connects data assets, the people who work with them, the policies that govern them, and the history of every decision made about them. Unlike a static diagram, it updates in real time as data environments change. Think of it as the memory layer for your data program. It records not just what your data is, but what has happened to it and why.
How is a context graph different from a knowledge graph?
Permalink to “How is a context graph different from a knowledge graph?”Knowledge graphs map entities and their relationships at a point in time. Context graphs add two things knowledge graphs lack: temporal context (every relationship has a valid_from and valid_to timestamp) and decision traces (relationships are first-class objects that carry the history of who created or modified them). A knowledge graph answers “what is related to what.” A context graph answers “why it changed and when.” See context graph vs. knowledge graph for the full architectural comparison.
What is a context graph in data governance?
Permalink to “What is a context graph in data governance?”In data governance, a context graph is the audit trail layer that connects data assets to the policies, decisions, and people that govern them over time. It answers governance questions that knowledge graphs cannot: “Was this dataset certified before it was used in this model training run?” “Who changed the data classification for this column, and when?” “Which governance policy is blocking this downstream report?” Context graphs make governance traceable, automatable, and queryable.
Can AI agents use context graphs?
Permalink to “Can AI agents use context graphs?”AI agents use context graphs as their persistent memory and reasoning substrate. When an agent needs to decide whether a dataset is trustworthy, a context graph provides the certification history, data quality incident record, and governance policy status with timestamps. When an agent makes a decision, the context graph records that decision as a new trace, creating the feedback loop that makes agents consistent across sessions. The Model Context Protocol (MCP) defines the transport layer; context graphs define the data structure that MCP-connected systems expose to agents.
What are decision traces in a context graph?
Permalink to “What are decision traces in a context graph?”Decision traces are the records of every governance action, certification event, and operational decision made about a data asset. In a context graph, each decision trace is a first-class object. It carries its own attributes: the actor who made the decision, the timestamp, the authority under which it was made, and any supporting evidence. Decision traces are what allow a context graph to answer “why” questions, not just “what” questions. Foundation Capital’s December 2025 thesis identified decision traces as the core primitive distinguishing context graphs from prior graph architectures.
What is reification in a context graph?
Permalink to “What is reification in a context graph?”Reification treats a relationship as a node in its own right, rather than only as an edge between two nodes. In a standard graph, “Column A is certified by Steward B” is an edge. With reification, that relationship becomes a node carrying its own attributes: when the certification was granted, under what policy, and whether it is currently valid. Reification enables temporal queries against relationships and is foundational to context graph architecture.
What is the W3C Context Graphs Community Group?
Permalink to “What is the W3C Context Graphs Community Group?”The W3C Context Graphs Community Group launched on February 24, 2026. It started with 56 members. Ron Itelman chairs the group. Its charter covers standardizing context graph interchange formats and query semantics across implementations. The W3C community group process is the same mechanism through which JSON-LD and SPARQL matured into foundational knowledge graph standards. The group’s formation signals that context graphs have reached the maturity level where interoperability standards are both needed and achievable.
What did Gartner say about context graphs?
Permalink to “What did Gartner say about context graphs?”Gartner’s Emerging Tech: Context Graphs Shape the AI Vendor Landscape report projects that over 50% of enterprise AI agent systems will incorporate context graphs by 2028. The projection is specific to enterprise deployments requiring governance, explainability, and audit — not general-purpose LLM applications. Separately, Gartner predicts 50% of AI agent deployment failures by 2030 will trace back to insufficient governance platform enforcement. Both signals position context graphs as approaching mainstream adoption for governance-heavy enterprise deployments.
How long does it take to build a context graph?
Permalink to “How long does it take to build a context graph?”For teams already running an active metadata platform with connected data sources, the context graph begins building from existing event streams immediately — no manual graph construction required. The timeline depends on your metadata maturity: organizations with active metadata and pre-built connectors can see initial graph structure within weeks. Teams starting from passive, manually-curated metadata face a longer ramp because the event stream that feeds the graph does not yet exist. See the implementation phases table above for a realistic phase-by-phase timeline.
Is a context graph the same as an ontology?
Permalink to “Is a context graph the same as an ontology?”No. An ontology defines formal concept taxonomy and logical rules. It specifies what concepts exist and how they relate in principle. It does not change over time and has no memory of past states. A context graph uses ontological definitions as its schema layer but adds temporal context, decision traces, and governance embedding. The ontology defines what relationship types are valid; the context graph records actual instances of those relationships with full history. They are complementary: ontologies define meaning, context graphs track operational history.
What companies are building with context graphs?
Permalink to “What companies are building with context graphs?”Interloom raised $16.5M in seed funding in March 2026 to build context graph infrastructure for regulated industries, with early customers including Commerzbank, Zurich Insurance, and Volkswagen. Atlan implements context graphs as a context layer for active data governance. Foundation Capital backed context graph infrastructure in December 2025 as part of a thesis on AI agent memory. The W3C Context Graphs Community Group includes 56 members spanning enterprise data teams, platform vendors, and AI infrastructure companies.
Why decision traces are the layer your data catalog is missing
Permalink to “Why decision traces are the layer your data catalog is missing”A context graph is the architecture that converts a passive data catalog into an operational memory system. Knowledge graphs and ontologies are useful layers — they map relationships and define meaning. But they carry no record of how those relationships changed, under what authority, or when. Context graphs add that history: the decision traces and governance embedding that make data trustworthy for AI agents and auditable for regulators.
Interloom’s $16.5M seed, the W3C community group formation, Foundation Capital’s trillion-dollar thesis, and Gartner’s 50%-by-2028 projection all landed in the same four-month window. That is not hype cycle noise. That is a technology category forming in real time. Verdantix is correct that context graphs are transitional — but the window for organizations to build this infrastructure before it is a requirement is measured in months, not years.
Data engineering leaders evaluating context graphs should start with a concrete question: does your current data catalog record why a dataset was certified, who approved the certification, and whether that certification is still valid? If the answer is no, your data infrastructure lacks the decision trace layer that AI agents and governance programs require. Context graphs are the architectural response to that gap.
For organizations on active metadata platforms, the context graph begins building from existing event streams. The infrastructure investment is lower than starting from scratch. The alternative is waiting for full standards maturity before building, which carries the risk of arriving late to an architecture that AI-dependent competitors are already running in production.
Share this article
