Context Graphs for AI Agents
Curated resource guide · Last updated February 2026
Context graphs are a structured data layer that captures decision traces, entity relationships, and temporal context for AI agents in production. This page curates the resources that matter — and separates the signal from the hype.
What you’ll find here
Context graphs have moved from VC thesis to active infrastructure debate in under three months. Foundation Capital called them AI’s trillion-dollar opportunity. Dharmesh Shah offered a reality check. FlexRule called the whole approach “decision archaeology”.
This guide curates the most substantive resources on both sides — foundational reading, technical deep dives, open source tools, critical perspectives, and industry-specific applications. Every entry is annotated so you know what you’re getting before you click.
What context graphs actually solve
Before diving into resources, a clear framing. Context graphs solve three foundational problems for AI agents in production:
Decision provenance
Every agent action is traceable to its input context, enabling post-hoc auditing and debugging.
Execution validation
Before an agent takes an action, the context graph enforces constraints — policy, budget, permissions.
Temporal context
Unlike static knowledge graphs, context graphs capture the history of decisions — allowing agents to learn from precedent.
What context graphs don’t solve
No architectural pattern fixes everything. Intellectual honesty matters here:
Real-time inference speed
Context graphs add validation latency. They’re not suited for sub-millisecond decision loops.
Bias elimination
As FlexRule points out, treating precedents as policy can reinforce existing biases without explicit governance rules.
Autonomous reasoning
Context graphs validate decisions — they don’t make them. The LLM still reasons; the graph enforces guardrails.
Context graph vs knowledge graph vs vector database
These are complementary tools, not competing ones. Most production systems combine two or three.
| Dimension | Context Graph | Knowledge Graph | Vector DB |
|---|---|---|---|
| Primary purpose | Decision traces, execution validation, temporal reasoning | Static entity relationships and domain modeling | Semantic similarity search over unstructured content |
| Temporal awareness | Built-in: valid_at / invalid_at timestamps on all edges | Limited: typically no native time modeling | None: embeddings are point-in-time snapshots |
| Query pattern | Graph traversal + temporal filters + policy constraints | Graph traversal (SPARQL, Cypher) | k-NN similarity search |
| Auditability | Full provenance: who decided what, when, and why | Partial: entity lineage only | None: no decision trace capability |
| Schema requirement | Intentional design: entities, relationships, and provenance rules | Ontology or schema definition | Schema-free: embed and index |
| Best used for | Agent decision validation in regulated, high-stakes workflows | Domain modeling, entity disambiguation, content linking | Content retrieval, semantic search, RAG pipelines |
| Latency profile | Medium: adds validation overhead (100–500ms typical) | Low to medium depending on graph depth | Low: sub-50ms for most queries |
The resources
12 resources organized by function. Each entry is annotated — you’ll know what you’re getting before you click.
Foundational Reading
The pieces that defined the context graph conversation — required reading before going deeper.
Foundation Capital — Jaya Gupta & Ashu Garg
The thesis that launched the context graph conversation. Argues the next trillion-dollar platforms won't add AI to existing systems of record — they'll capture the decision traces that make data actionable. Defines a context graph as "a living record of decision traces stitched across entities and time so precedent becomes searchable."
Dharmesh Shah — Dharmesh Shah (HubSpot)
Shah finds the context graph idea intellectually compelling — but offers a reality check. Most companies still struggle with basic data unification. Expecting them to deploy sophisticated "decision lineage" may be premature. A grounded counterbalance to the VC enthusiasm.
Constellation Research — Esteban Kolsky
Analyst-grade perspective from Constellation Research. Argues most enterprise AI failures aren't caused by bad models — they're caused by teams scaling automation on search-era infrastructure. Positions context graphs as the missing architectural layer.
Glean — Arvind Jain
Glean's CEO explains how context graphs emerge from connecting existing enterprise systems — connectors, indexes, graphs, and memory — into an entirely new data platform designed for agentic automation, not reporting.
Technical Deep Dives
What context graphs actually solve at the implementation level — architecture patterns, decision runtimes, and the engineering tradeoffs involved.
The Context Graph — Patrick Joubert
Breaks down the core definition of a context graph — how it differs from general-purpose knowledge graphs, what makes it temporal, and why decision state is the missing layer for production AI agents.
The Context Graph — Patrick Joubert
A structural comparison: knowledge graphs represent static domain relationships; context graphs add temporal validity, execution traces, and decision provenance. Different tools for different jobs.
Rippletide
Covers the three context failures behind AI agent unreliability — applicability, temporal validity, and traceability — and how context graphs address each one structurally. Also honest about what context graphs don't solve.
CloudRaft
The most comprehensive hands-on guide available. Covers database selection (FalkorDB, Neo4j), entity modeling, hybrid architectures combining vectors and graphs, and a realistic timeline: 4–8 weeks for proof of concept, 3–6 months for production.
TrustGraph — Daniel Davis
A specification-level document arguing that context graphs should be triples-based (subject-predicate-object) and optimized for explainable AI. Traces the evolution from basic RAG to GraphRAG to ontology-aware context graphs.
Tools & Open-Source Frameworks
The building blocks — open-source projects and platforms shipping context graph infrastructure today.
Zep (Open Source)
Open-source framework for temporally-aware knowledge graphs. Supports incremental updates, bi-temporal data models, and hybrid search (semantic + BM25 + graph traversal). Powers Zep's context engine. Benchmarks: 94.8% on Deep Memory Retrieval, P95 latency of 300ms.
Graphlit — Kirk Marple
Graphlit argues you can't capture decision traces without first solving operational context — who owns what, how entities relate, what changed. Proposes CRM as the "entity spine" for resolving multimodal content into a coherent graph.
Critical Perspectives
Not everyone agrees. These pieces challenge the context graph thesis — essential reading for a balanced view.
FlexRule
The strongest critique in the space. FlexRule argues context graphs are "decision archaeology, not decision intelligence" — they address symptoms by collecting traces after execution rather than modeling decisions explicitly before execution. Proposes DMN-based decision modeling as the alternative. Worth reading for the architectural counter-argument alone.
Industry & Vertical Applications
Where context graphs meet regulated industries, compliance requirements, and sector-specific constraints.
WeBuild-AI
Maps context graphs to specific compliance demands: FCA Consumer Duty in financial services, safety case documentation in energy, and supply chain traceability in manufacturing. Honest about the timeline — full implementation is a 2026–2027 milestone for most enterprises.
Latent Space — swyx
Frames context graphs through the lens of coding agents and open standards. Covers Agent Trace — a cross-company initiative (Cursor, Vercel, Cloudflare) for mapping code-to-context. The first domain-specific context graph specification to gain multi-vendor agreement.
Frequently asked questions
What is a context graph for AI agents?
A context graph is a structured data layer that captures decision traces, entity relationships, and temporal context for AI agents in production. Unlike static knowledge graphs, context graphs record not just what happened but why — including exceptions, overrides, and precedents — enabling agents to reason over real organizational history rather than isolated data chunks.
How is a context graph different from a knowledge graph?
Knowledge graphs represent static domain relationships (e.g. "Company X is in Industry Y"). Context graphs extend this by adding temporal validity, decision provenance, and execution traces. A knowledge graph tells you what exists; a context graph tells you what happened, why it was decided, and whether the decision still applies.
What problems do context graphs solve for AI agents?
Context graphs solve three foundational problems: (1) Decision provenance — every agent action is traceable to its input context. (2) Execution validation — constraints like policies, budgets, and permissions are enforced before an agent acts. (3) Temporal context — agents can reason over the history of decisions, not just the current state.
What are the limitations of context graphs?
Context graphs add latency and are not suited for sub-millisecond decision loops. They require intentional knowledge engineering and upfront schema design. They validate decisions but do not make them. And treating historical precedents as policy can reinforce existing biases if not paired with explicit governance rules.
How do context graphs compare to vector databases and RAG?
Vector databases excel at semantic similarity search over unstructured content. RAG retrieves text chunks to augment LLM prompts. Context graphs capture structured relationships and decision traces over time. In practice, production systems often combine all three: vectors for content retrieval, RAG for prompt augmentation, and context graphs for relationship reasoning and auditability.
Building with context graphs in production? We’re tracking the structural patterns behind agent reliability — what breaks, what holds, and what infrastructure actually fixes it.