Context Graphs — Agent Trajectories and Organizational Memory

Synthesis of Foundation Capital’s (Ashu Garg) context graphs concept: an infrastructure layer that turns agent traces into a compounding asset.

Definition

Context Graph: a graph that captures AI agent decision-making trajectories. Not just what the agent did, but how and why. A system of records for decisions, not just data.

“A system of record for decisions, not just data.” — Dharmesh Shah (HubSpot)

“We’ve entered the era of context, where organizational knowledge becomes the competitive differentiator.” — Aaron Levie (Box)

Key Concepts

Agent Trajectories

When an agent performs a task, it traverses the organization’s state space: touches systems, reads data, calls APIs. This trajectory is a decision trace.

Each trajectory captures:

• Reasoning steps and intermediate decisions
• Tool selections and API calls
• Error recovery and backtracking paths
• Context state at each decision point

As trajectories accumulate, an organizational world model emerges.

Decision Traces

Difference between “traces” in observability and “decision traces”:

• Observability trace = what happened (latency, errors, spans)
• Decision trace = why it happened (reasoning, alternatives, context)

Agents sit on a unique set of trajectories representing real decision-making in organizations, something otherwise impossible to observe directly.

Half-life of Decisions

Decisions have an expiry date:

• Policies change, teams restructure
• Agents need to know which precedents are still relevant, not just which ones exist
• Requires a mechanism for decay and re-evaluation

Connection to Process Mining

Context graphs extend Process Mining ideas (Celonis, UiPath):

Practical Cycle: Capture -> Retrieve -> Apply

Not a formal standard, but a working pattern for using decision traces:

1. Capture: record agent decisions into a graph (what + why)
2. Retrieve: before new task, search for similar precedents
3. Apply: adapt found patterns to current situation

Each successful action improves future ones, creating compound learning.

Note: “Context graphs” is Foundation Capital’s VC thesis term, not a technical standard. The pattern matters, not the branding.

Ontology Debates

Three approaches to structuring context graphs:

Emergent Ontology (PlayerZero / Animesh)

• No prescribed structure upfront
• Agents discover the ontology through usage
• Decision traces compile into organizational world models
• Ontology is found, not declared

Prescriptive Ontology (Palantir)

Three layers:

1. Semantic layer: objects and relationships
2. Kinetic layer: actions and flows
3. Dynamic layer: simulations and decision logic

Hybrid (Graphlit et al.)

• Core entities (Person, Organization, Account, Event) are stable, no need to discover
• Domain relationships and decision patterns: let agents discover
• Pragmatic balance: skip discovery for the obvious

Trillion-Dollar Thesis

Economic value of context graphs:

1. Reduced compute waste: reuse proven decision paths instead of re-deriving them
2. Faster onboarding: new agents bootstrap from accumulated knowledge
3. Compound learning: every action improves future ones
4. Enterprise memory: organization retains agent expertise across team changes
5. Network effects: value grows with graph enrichment (more traces = better precedents)

First mover to capture decision traces in a high-value domain builds a compounding moat.

Practical Implementation

For Small Projects and Assistants

Context graphs don’t require enterprise scale. Already working at small scale:

1. Session history (already implemented in solograph)

   • session_search: search past Claude Code sessions
   • Each session = decision trajectory

2. CodeGraph (already implemented)

   • codegraph_query: Cypher queries on code graph
   • codegraph_explain: architectural overview
   • codegraph_shared: shared packages across projects
   • Structural memory about code

3. Knowledge Base (already implemented)

   • kb_search: semantic search
   • Decisions, patterns, precedents

4. Source Graph (already implemented)

   • source_search: Telegram, YouTube
   • source_related: related videos by tags
   • External knowledge sources in graph

What to Strengthen

• Decision logging: capture in KB not just results, but why a decision was made
• Agent trajectory export: export trajectories from session_search in structured format
• Pattern mining: find recurring decision patterns across sessions
• Precedent retrieval: before new task, auto-search for similar past solutions
• Decay mechanism: mark stale decisions (half-life)

Connection to Harness Engineering

Context graphs and harness engineering complement each other:

Together: harness constrains the agent, context graph helps choose the best path within constraints.

Evaluation via Context Graphs

Assess agents through their decision traces:

• Did the agent traverse expected graph edges?
• Are constraints adequate?
• Do similar contexts produce consistent trajectories?
• Is decision quality degrading over time?

Industry and Adoption

• Dharmesh Shah (HubSpot): “system of record for decisions”
• Aaron Levie (Box): “era of context”
• Arvind Jain (Glean): “the concept finally has a name”
• Cognition (Devin): integrated into product
• Fortune 500 CIOs: requesting roadmap
• OpenAI + Anthropic: employees sharing as required reading

Sources:

• Foundation Capital — Context Graphs: AI’s Trillion-Dollar Opportunity (Ashu Garg)
• Foundation Capital — Context Graphs, One Month In (Jan 2026)
• Foundation Capital — Why Context Graphs Are the Missing Layer
• Graphlit — Context Graphs: What the Ontology Debate Gets Wrong