Kicker: Foundation Capital's December 2025 essay reframed enterprise AI's missing piece. Six months in, the term is everywhere — but the buyer-side category is still being assembled.

Context Graphs: The Emerging Enterprise Memory Layer for Agentic AI

In December 2025, Jaya Gupta and Ashu Garg of Foundation Capital published an essay arguing that the next trillion-dollar enterprise software opportunity is neither a model nor an agent framework, but a substrate underneath them: the context graph. By their definition, a context graph is a living record of decision traces — who decided what, with what information, under what constraints, and why — stitched across entities and time so precedent becomes searchable. Where Salesforce and SAP became systems of record, the authors argue, context graphs will become the systems of decisions that agentic AI requires.

Six months later the term is in Gartner research, a16z's "Big Ideas 2026" thesis, Aaron Levie's TechCrunch interviews, Atlassian's Team '26 keynote, and Neo4j's "Context Gap" thesis. Vendors from agent-memory startups to knowledge graph incumbents to enterprise search platforms are now relabeling. For CIOs being pitched on a "context layer," "memory fabric," "work graph," or "enterprise ontology," the practical question is the one no slide deck can answer: is this a category to procure, a capability to build, or a thesis to revisit in 2027?

This article defines the concept honestly, maps the four vendor archetypes claiming the space, surfaces the skeptics, and lays out the open questions every CIO should bring to a context-graph conversation.

:::stat-row Months since the term entered the discourse | 6 (Dec 2025 — May 2026) Gartner forecast — AI agent systems leveraging context graphs by 2028 | >50% Foundation Capital essays in the canon | 4 (incl. Aaron Levie co-byline) Pure-play startups explicitly using the term | ~10 (mostly seed / Series A) :::

What a Context Graph Actually Is

The simplest way to understand the concept is by what it is not. A knowledge graph captures facts and entities — your customers, products, contracts, and the relationships between them. A vector database captures semantic similarity for retrieval. A conversational memory layer like Mem0 or Letta captures what a single agent has been told by a single user. A data catalog captures metadata about where information lives.

A context graph, as the Foundation Capital authors define it, captures something none of those do: the trace of decisions an organization actually made. Why was the discount approved on this deal but not that one? Which exception did finance allow last quarter when the same situation arose? What did the credit committee weigh when it approved the loan despite the policy guidance? Three properties distinguish the construct:

• Decision traces, not just facts. Nodes represent choices and the reasoning behind them, not just entities.
• Temporal by design. Traces are stamped with time so the graph supports "show me how we used to handle this" queries, not just "what is true now."
• Emergent ontology. Rather than modeling the schema upfront, the graph accretes structure from agent execution traces — what Foundation Capital calls a "living record."

The third property is the most contentious and the most consequential for buyers. Traditional knowledge-graph projects fail when the ontology committee outlasts the executive sponsor. Context graphs propose to skip that step: let agents act, capture the traces, mine the patterns. Whether that works at enterprise scale is one of the open questions discussed below.

"A context graph is a living record of decision traces, stitched across entities and time, so precedent becomes searchable." — Foundation Capital framing, December 2025

Why the Existing Layers Are Insufficient — The Argument

The thesis assumes today's stack leaves a gap. The argument runs:

Existing layer	What it does well	Where it falls short for agents
LLM context window	Carries short-term task state	Forgets at session boundary; no cross-agent shared memory
Vector RAG / GraphRAG	Retrieves relevant facts for grounding	Improves what an agent knows, not what it should do in process
Knowledge graph	Models entities and relationships	Static facts; no decision history; ontology overhead
Conversational memory (Mem0, Letta)	Personalizes a single agent for a single user	Single-tenant; not an organization-wide substrate
Data catalog / lineage	Tracks where data lives and how it flows	Describes pipelines, not decisions
Workflow / BPM engine	Encodes the prescribed process	Captures the doc, not the actual exception-laden practice

Every layer above is real and used. The Foundation Capital argument is that none of them, alone or stacked, gives a fleet of agents what an experienced human employee carries: a sense of how this organization decides things, what has been tried, what worked, what did not, and why. Whether that gap warrants a new product category — or is simply a feature roadmap for the layers that already exist — is what the next sections wrestle with.

The Four Archetypes Claiming the Category

A buyer asked to "evaluate context graph platforms" today is being asked to compare radically different things. Roughly four archetypes have emerged.

Archetype 1 — Agent-Memory SDKs

Lightweight memory libraries developers add to an agent runtime. Often open-source-led, sold to AI engineering teams, priced like developer infrastructure.

• Zep / Graphiti — Temporal knowledge graph for agents; the open-source Graphiti core has the strongest published scores on long-term memory benchmarks like LongMemEval.
• Mem0 — Personalization-oriented memory with user / agent / run / app scopes; YC alum, Series A.
• Letta (formerly MemGPT) — Self-editing agent runtime memory popularized by the MemGPT paper.
• Cognee — Document-to-knowledge-graph memory for agents.
• Supermemory — VPC-deployable memory infrastructure.
• Graphlit, TrustGraph, LangMem — Adjacent OSS / commercial primitives, several of which use the "context layer" framing explicitly.

Buyer profile: AI platform engineering teams, not procurement. Deals are SDK-scale, not eight-figure software contracts.

Archetype 2 — Knowledge-Graph Databases Retrofitting GraphRAG

The graph-database incumbents argue they have been the context layer all along. Their pivot toward agentic AI is the most aggressive of the four archetypes.

• Neo4j — ~$581M raised, ~$2B valuation, $100M+ committed to agentic-AI investment in 2025; "Agentic GraphRAG" headlining its NODES 2026 conference; the Futurum Group coverage of Neo4j's "Context Gap" thesis is the most-cited incumbent answer to Foundation Capital.
• TigerGraph — GraphRAG plus added vector search.
• Stardog — Semantic-web / ontology-first; strong in regulated industries.
• Ontotext (GraphDB), Memgraph, ArangoDB, AWS Neptune, SAP Datasphere Knowledge Graph — All positioning into agentic GraphRAG.

Buyer profile: Existing graph-database customers, data architecture teams, regulated industries that need provenance.

Archetype 3 — Enterprise Work-Graph and Ontology Platforms

The most credible end-to-end answers, because they already operate at enterprise scale and already capture some form of cross-application context.

• Palantir AIP / Ontology — Arguably already shipped a proprietary context graph; Ontology explicitly markets "decisions, not just data." The platform several analysts hold up as evidence that the category is real but largely Palantir's to defend.
• Glean — $7.2B valuation (Series F, June 2025), $200M+ ARR; the "Work Graph" plus Glean Agents, with reported 100M+ agent actions per year.
• Atlassian Rovo / Teamwork Graph — At Team '26 (May 2026) Atlassian opened the Teamwork Graph (~150B+ connections) to third-party agents via Model Context Protocol; ~90% of enterprise cloud customers reported using Rovo.
• Microsoft Graph in Fabric, Microsoft 365 Copilot, Project GraphRAG — Microsoft's existing graph plus Fabric is the default context substrate for Microsoft-shop enterprises.
• Box AI Hubs — Aaron Levie's content-side context graph thesis; Levie co-authored the Foundation Capital "case for context graphs" essay.
• Writer Knowledge Graph — Enterprise knowledge graph integrated with Writer's domain-specific LLMs.

Buyer profile: CIOs, CDOs, and digital workplace owners. These are the only archetype-3 vendors with eight-figure enterprise deals already closing today.

Archetype 4 — Pure-Play Context-Graph Startups

The names that explicitly market themselves under the new label, often citing the Foundation Capital essays directly.

• Interloom — Munich-based; $16.5M seed (DN Capital, March 2026). Markets explicitly as "enterprise memory for AI agents" and ingests tickets, emails, and transcripts to build a context graph. Disclosed customers include Zurich Insurance, JLL, Fiege, Commerzbank, and Volkswagen.
• Foundation Capital portfolio examples — PlayerZero (production engineering, with a public "Context Graphs" page), Maximor (finance), Regie (demand gen), Tessera, Tonkean. These are systems-of-agents plays whose execution traces accrete a context graph as a side effect rather than the primary product.
• IndyKite, Tekst, Amnic, Ema.ai — Earlier-stage entrants that adopt the context-graph or context-layer label explicitly.

Buyer profile: Innovation budgets, AI office pilots, function-specific use cases (CFO, customer ops, support).

A side note on who is not yet in the category: as of May 2026, no major foundation-model lab (OpenAI, Anthropic, Google) has shipped a branded "context graph" product. Anthropic's Model Context Protocol and OpenAI's Responses API memory are enabling primitives, not the layer itself.

The Skeptical View

The CIO audience deserves the dissent. The most credible skeptics in early 2026:

• Sanjeev Mohan (formerly Gartner, now SanjMo) has compared context graphs to the data mesh debate of 2020–2022 — a thesis that produced a great deal of conference content, no shared standard, and limited durable adoption. He has warned the industry will "squander 2026 debating context graphs."
• Latent.Space has written that the thesis "promises a whole lot but is not very prescriptive," and that the prescriptive vs. emergent ontology debate inside Foundation Capital's own follow-ups is unresolved.
• HubSpot CTO Dharmesh Shah has noted publicly that most enterprises still struggle with basic data unification, making the leap to decision-trace memory aspirational for the median Fortune 1000 organization.

The pattern these critics identify is real: the term outran the standards. There is no equivalent of the SQL standard, no shared schema for a decision trace, no Magic Quadrant, and limited public evidence of named, large-enterprise deployments outside the archetype-3 incumbents (Palantir, Glean, Atlassian, Microsoft) and Foundation Capital's portfolio reference accounts.

The CIO's Open Questions

For a CIO sitting through a vendor pitch, these are the questions that separate the construct from the marketing.

1. What unit of capture does this product treat as primary? A decision trace, an entity, a document chunk, a conversation, or a metadata record? The answer reveals which archetype you are actually buying — and whether it solves the problem you are trying to solve.

2. Ontology — prescribed or emergent? Foundation Capital's own follow-up posts argue for emergent ontology mined from agent traces. Knowledge-graph incumbents argue for prescribed ontologies modeled by domain experts. This is not a small difference; it determines time-to-value, governance burden, and whether the project survives the first reorg.

3. Where does it sit relative to your data catalog, MDM, and identity graph? If the answer is "alongside," you have just added a fifth governance surface. If the answer is "on top of," demand the integration story.

4. Who owns it? The most under-asked question. Is the context graph a CDO asset, a CIO platform, a head-of-AI initiative, or a platform-engineering deliverable? The four archetypes correspond loosely to four different buyers; choosing one archetype without resolving ownership creates a budget orphan.

5. Build vs. buy. For most enterprises in 2026 the realistic answer is buy a work-graph platform you already use (Microsoft 365 + Fabric, Atlassian Rovo, Glean, Palantir Ontology) and integrate an agent-memory SDK (Zep, Mem0, Letta) inside the agent runtime. Pure-play context-graph products are credible for innovation budgets and bounded use cases, not yet for cross-enterprise standardization.

6. What is the exit strategy? Given the category is six months old and consolidation is likely, demand exportable schemas, open formats (RDF, OpenCypher, MCP), and contractual commitments to data portability. This will not be the only context graph the enterprise ever owns.

A Pragmatic Stance for 2026

The honest summary: the concept is real and the category is not yet shippable as a single procurement. CIOs have four reasonable moves:

• Don't buy a "context graph" as a category. Buy a work-graph platform you can already justify on adjacent value (search, agents, productivity), and let it accrete decision traces as a byproduct.
• Treat agent-memory SDKs as developer infrastructure. Standardize on one for the AI platform, not as a strategic procurement.
• Pilot one pure-play in one function where decision traceability is the clear value (finance close, support escalation, deal desk). Treat it as a learning investment, not a platform commitment.
• Revisit the category in Q1 2027, after Gartner's first formal coverage, named Fortune 500 disclosures from the pure-plays, and consolidation among the agent-memory SDK vendors.

The single biggest risk is buying the brochure. The single biggest opportunity is asking, in every agent rollout you do this year, where do these decision traces go, and who can query them in five years? That question alone tells you whether your enterprise is building a context graph by accident, or paying a vendor to do it on purpose.

Frequently Asked Questions

Q: Is a context graph the same as a knowledge graph? A: No. A knowledge graph captures entities and their relationships as facts. A context graph layers in decision traces — who decided what, with what information, when, and why — and treats time as a first-class dimension. A knowledge graph answers "what is true"; a context graph answers "how does this organization decide."

Q: Do we need a context graph if we already have GraphRAG and a vector database? A: Probably not as a separate purchase, yet. GraphRAG and vector retrieval improve what an agent knows. The unsolved problem is what an agent should do in process, especially on edge cases — and whether that judgment is captured for future agents. If your agent fleet is small and supervised, a vector + GraphRAG stack is sufficient. If you are scaling to dozens of agents acting across functions, decision-trace memory becomes more interesting.

Q: Who in the enterprise should own the context graph? A: The category is too new for a definitive answer, but the credible options are: (1) the CDO, if you treat it as a data asset; (2) the CIO platform organization, if you treat it as shared infrastructure; (3) the head of AI / AI Office, if you treat it as agent-runtime tooling. The wrong answer is "no one" or "all of the above."

Q: Should we wait for Gartner to define the category before acting? A: For procurement standards, yes — expect a Magic Quadrant equivalent in 2027 at the earliest. For experimentation, no — every agent rollout this year is generating decision traces whether you capture them or not. The cost of capturing-and-deferring is low; the cost of discarding two years of traces and starting over later is high.

Q: Is Palantir already a context graph platform? A: Several analysts argue yes, in everything but the marketing label. Palantir AIP / Ontology has long modeled decisions, not just data, and operates at enterprise scale. The trade-off is the well-understood Palantir trade-off: depth, opinionatedness, and lock-in. Whether that is the right answer depends on industry, regulatory posture, and tolerance for a sole-source platform commitment.

Q: How does this relate to the Model Context Protocol (MCP)? A: MCP is a wire protocol for connecting agents to tools and data sources, including context stores. It is an enabling primitive, not a context graph product. Atlassian opening Teamwork Graph to third-party agents over MCP at Team '26 is an example of the protocol carrying context-graph traffic; the protocol does not define what a context graph contains.

Sources & Further Reading

• Jaya Gupta and Ashu Garg, "AI's trillion-dollar opportunity: Context graphs," Foundation Capital, December 22, 2025.
• Foundation Capital, "Context graphs, one month in" (February 2026) and "The case for context graphs" with Aaron Levie (February 2026); "Why context graphs are the missing layer for AI."
• Andreessen Horowitz, "Big Ideas 2026: The Enterprise Orchestration Layer."
• Atlan, "Gartner on Context Graphs" (March 2026).
• Metadata Weekly (Substack), critique of the context-graph thesis.
• Latent.Space, "Context graphs — hype or actually?"
• Diginomica, "Context graphs unlock a new seam of enterprise knowledge for AI agents."
• Futurum Group, "Does Neo4j's Context Gap thesis expose enterprise AI's biggest blind spot?"
• TrustGraph, "Context Graph vs. Knowledge Graph."
• Arize, "How context graphs turn agent traces into durable business assets."
• Fortune, "Interloom raises $16M for AI agent memory" (March 2026).
• Glean Series F announcement (June 2025); Atlassian Team '26 keynote coverage (May 2026); Palantir Ontology product page.

---

Related on CIOPages: Agentic AI: What CIOs Need to Know · Generative AI Operating Models · The Enterprise of Agents · Buyer's Guide: AI Agent & Agentic AI Platforms · Buyer's Guide: Vector Databases & AI Search