Skip to main content

Graph Capabilities: Built-In Graph Traversal

Info

Last Updated: 2026-01-09 | Version: v0.29.0+

Navigate relationships and traverse knowledge graphs using Cortex's document-oriented architecture.

Note: This document covers advanced graph patterns and performance analysis. For basic context chain operations (creating, updating, querying), see Context Chains.

Overview​

Cortex is built on Convex, a document-oriented database, but provides graph-like querying through references and relationships. We call this Graph-Lite - you get graph capabilities without running a separate graph database.

Your data IS a graph - it's just stored in documents instead of native graph nodes and edges.

The Implicit Graph​

Nodes (Entities)​

Every entity in Cortex is a graph node:

EntityDescription
AgentsAI agents, human operators
UsersEnd users, customers
ContextsWorkflow tasks, hierarchies
ConversationsMessage threads
MemoriesAgent knowledge
FactsExtracted knowledge (optional)

Edges (Relationships)​

References between documents are graph edges:

conversationRef → Links Memory to Conversation
immutableRef    → Links Memory to Immutable Record (Fact, KB Article)
mutableRef      → Links Memory to Live Data
parentId        → Links Context to Parent Context
userId          → Links anything to User
memorySpaceId   → Links Memory/Context/Conversation to Memory Space
contextId       → Links Memory to Workflow Context
fromAgent/toAgent → Links A2A Messages between Agents

Visualizing the Graph​

Cortex Implicit Graph Structure
User-123
Root entity with connections to conversations and memories
Conversation-456
ACID layer — triggered Context-001
Context-001 (Root)
Parent of Finance and Customer contexts
Memory-abc
Vector layer — references Conversation, Fact, and Context
Key Relationships

Context hierarchy: Context-001 → Context-002 (Finance) → finance-agent
Memory links: Memory → Conversation, Fact, Context
A2A messaging: finance-agent → customer-agent (dual-write)

Built-In Graph Traversals​

1. Context Chain Navigation​

Context chains provide hierarchical graph traversal:

// Get complete context chain (multi-hop graph walk)
const chain = await cortex.contexts.get("ctx-child", {
  includeChain: true, // ← Graph traversal!
});

console.log("Graph nodes accessed:");
console.log("- Current:", chain.current.purpose);
console.log("- Parent:", chain.parent?.purpose); // 1-hop up
console.log("- Root:", chain.root.purpose); // N-hops to root
console.log("- Siblings:", chain.siblings.length); // Lateral traversal
console.log("- Children:", chain.children.length); // 1-hop down
console.log("- Ancestors:", chain.ancestors.length); // All hops to root
console.log("Total nodes in graph:", chain.totalNodes);

Equivalent Graph Query:

// Neo4j Cypher equivalent
MATCH (current:Context {id: $contextId})
OPTIONAL MATCH (current)-[:CHILD_OF]->(parent:Context)
OPTIONAL MATCH (parent)-[:CHILD_OF*]->(root:Context)
OPTIONAL MATCH (current)<-[:CHILD_OF]-(siblings:Context)
OPTIONAL MATCH (current)-[:CHILD_OF]->(children:Context)
RETURN current, parent, root, siblings, children

Performance:

  • Depth 0-2: 50-100ms (single Convex query)
  • Depth 3-5: 100-300ms (2-3 Convex queries)
  • Depth 6+: 300ms+ (consider native graph DB)

2. Conversation Tracing​

Trace memories back to their source conversations:

// Memory → Conversation link via conversationRef
const memory = await cortex.vector.get("agent-1", "mem-abc123", {
  includeConversation: true, // ← Graph traversal!
});

console.log("Memory content:", memory.content);
console.log("Source conversation:", memory.conversationRef?.conversationId);

// Get full conversation from ACID layer
if (memory.conversationRef) {
  const conversation = await cortex.conversations.get(
    memory.conversationRef.conversationId,
  );
  console.log(
    "Source messages:",
    conversation.messages.map((m) => m.content),
  );
}

// Full audit trail preserved in ACID layer

Graph Structure:

Memory-abc (Vector Layer)
  │
  └──[conversationRef]──> Conversation-456 (ACID Layer)
                               │
                               └── Messages: [msg-001, msg-183, ...]

Use Cases:

  • Audit trails: "Where did this fact come from?"
  • Compliance: "Show original user statement for this memory"
  • Context enrichment: "Get full conversation around this fact"

3. Agent Communication Graph​

A2A communication creates an agent relationship graph:

// Get all A2A messages between two agents
const conversation = await cortex.a2a.getConversation(
  "finance-agent",
  "hr-agent",
  {
    since: new Date("2025-10-01"),
    minImportance: 70,
    tags: ["budget"],
  },
);

console.log(`${conversation.messageCount} messages exchanged`);
console.log("Communication graph:");
conversation.messages.forEach((msg) => {
  console.log(`  ${msg.from} ──[SENT_TO]──> ${msg.to}: ${msg.message}`);
});

Graph Structure:

Agent: finance-agent
  │
  ├──[SENT_TO {importance: 85}]──> Agent: hr-agent
  │
  ├──[SENT_TO {importance: 90}]──> Agent: legal-agent
  │
  └──[RECEIVED_FROM]──> Agent: ceo-agent

Multi-Hop Example:

// Find all memory spaces finance-agent has communicated with (1-hop)
const directConnections = await cortex.vector.search("finance-agent", "*", {
  sourceType: "a2a",
  limit: 100,
});

const connectedSpaces = new Set(
  directConnections
    .filter((m) => m.metadata?.toMemorySpace)
    .map((m) => m.metadata.toMemorySpace as string),
);

console.log("Direct collaborators:", Array.from(connectedSpaces));

// Find second-degree connections (2-hop)
const secondDegree = new Set();

for (const space of connectedSpaces) {
  const theirConnections = await cortex.vector.search(space, "*", {
    sourceType: "a2a",
    limit: 100,
  });

  theirConnections.forEach((m) => {
    const toSpace = m.metadata?.toMemorySpace as string;
    if (toSpace && toSpace !== "finance-agent") {
      secondDegree.add(toSpace);
    }
  });
}

console.log("2nd-degree collaborators:", Array.from(secondDegree));

Performance: 2-hop query = 100-200ms (acceptable for collaboration queries)

4. User Data Graph​

Trace all data related to a user across the entire system:

// User → Everything relationship
async function getUserDataGraph(userId: string) {
  const graph = {
    user: await cortex.users.get(userId),
    conversations: await cortex.conversations.list({ userId }),
    contexts: await cortex.contexts.list({ userId }),
    memories: [] as any[],
    facts: [] as any[]
  };

  // Get memories from all memory spaces
  const spaces = await cortex.memorySpaces.list();

  for (const space of spaces.spaces) {
    const spaceMemories = await cortex.vector.search(space.memorySpaceId, '*', {
      userId,
      limit: 100
    });

    graph.memories.push(...spaceMemories);
  }

  // Get facts about user
  const userFacts = await cortex.facts.list({
    memorySpaceId: "*", // Search all spaces
    userId
  });

  graph.facts.push(...userFacts);

  return graph;
}

// Result: Complete graph of user's data
{
  user: { id: 'user-123', data: { displayName: 'Alex' }, ... },
  conversations: [5 conversations],
  contexts: [12 workflows],
  memories: [234 memories across 8 memory spaces],
  facts: [45 extracted facts]
}

Graph Visualization:

User-123
  ├──> 5 Conversations
  ├──> 12 Contexts
  ├──> 234 Memories (across 8 agents)
  └──> 45 Facts

Use Case: GDPR data export, user analytics, personalization

Graph-Lite Query Patterns​

Pattern 1: Audit Trail Reconstruction​

Trace the complete path from user request to agent action:

async function traceAuditTrail(conversationId: string) {
  const trail = [];

  // Step 1: Get originating conversation
  const conversation = await cortex.conversations.get(conversationId);
  trail.push({ type: "conversation", entity: conversation });

  // Step 2: Find contexts triggered by this conversation
  const contexts = await cortex.contexts.search({
    "conversationRef.conversationId": conversationId,
  });

  for (const context of contexts) {
    trail.push({ type: "context", entity: context });

    // Step 3: Find all child contexts (workflow tree)
    const children = await cortex.contexts.getChildren(context.id, {
      recursive: true,
    });

    trail.push(...children.map((c) => ({ type: "context", entity: c })));

    // Step 4: Find memories linked to this context
    const contextMemories = await cortex.memory.search("*", "*", {
      metadata: { contextId: context.id },
    });

    trail.push(...contextMemories.map((m) => ({ type: "memory", entity: m })));
  }

  return trail;
}

// Usage
const trail = await traceAuditTrail("conv-456");

console.log("Audit trail (graph path):");
trail.forEach((node, i) => {
  console.log(`${i}. [${node.type}] ${node.entity.id}`);
});

// Output:
// 0. [conversation] conv-456
// 1. [context] ctx-001 (root)
// 2. [context] ctx-002 (child: finance approval)
// 3. [memory] mem-abc (finance agent's memory)
// 4. [context] ctx-003 (child: customer notification)
// 5. [memory] mem-def (customer agent's memory)

Performance: ~200-500ms for typical workflow (5-10 nodes)

Pattern 2: Knowledge Graph Traversal​

Find related facts via entities:

async function findRelatedFacts(factId: string) {
  // Get the source fact
  const fact = await cortex.immutable.get("fact", factId);

  if (!fact.data.entities || fact.data.entities.length === 0) {
    return [];
  }

  // Find other facts mentioning same entities
  const relatedFacts = [];

  for (const entity of fact.data.entities) {
    const facts = await cortex.immutable.search(entity, {
      type: "fact",
      limit: 10,
    });

    relatedFacts.push(...facts);
  }

  // Deduplicate
  const unique = Array.from(
    new Map(relatedFacts.map((f) => [f.id, f])).values(),
  ).filter((f) => f.id !== factId);

  return unique;
}

// Example
const fact = {
  id: "fact-1",
  fact: "Alice works at Acme Corp",
  entities: ["Alice", "Acme Corp"],
};

const related = await findRelatedFacts(fact.id);
// Returns:
// - "Alice manages the engineering team"
// - "Acme Corp is headquartered in San Francisco"
// - "Alice joined Acme Corp in 2020"

Graph Pattern (what we're simulating):

Fact-1 ──[MENTIONS]──> Entity: Alice <──[MENTIONED_BY]── Fact-2
Fact-1 ──[MENTIONS]──> Entity: Acme Corp <──[MENTIONED_BY]── Fact-3

Pattern 3: Workflow Dependency Graph​

Find all dependent contexts in a workflow:

async function getWorkflowDependencies(rootContextId: string) {
  const visited = new Set();
  const dependencies = [];

  async function traverse(contextId: string, depth = 0) {
    if (visited.has(contextId)) return;
    visited.add(contextId);

    const context = await cortex.contexts.get(contextId);
    dependencies.push({ context, depth });

    // Traverse children
    for (const childId of context.childIds) {
      await traverse(childId, depth + 1);
    }
  }

  await traverse(rootContextId);

  return dependencies;
}

// Usage
const deps = await getWorkflowDependencies("ctx-root");

console.log("Workflow graph:");
deps.forEach(({ context, depth }) => {
  const indent = "  ".repeat(depth);
  console.log(`${indent}${context.purpose} (${context.status})`);
});

// Output:
// Process refund request (active)
//   Approve refund (completed)
//   Send apology email (completed)
//   Update CRM (active)

Performance: O(n) where n = number of nodes, typically <500ms for <50 nodes

Pattern 4: Agent Expertise Network​

Build a graph of agent capabilities and collaborations:

async function buildMemorySpaceNetwork() {
  const spaces = await cortex.memorySpaces.list();
  const network = {
    nodes: [] as any[],
    edges: [] as any[]
  };

  // Add memory space nodes
  for (const space of spaces.spaces) {
    network.nodes.push({
      id: space.memorySpaceId,
      name: space.name || space.memorySpaceId,
      type: space.type,
      metadata: space.metadata
    });
  }

  // Find collaboration edges (via A2A)
  for (const space of spaces.spaces) {
    const collaborations = await cortex.vector.search(space.memorySpaceId, '*', {
      sourceType: 'a2a',
      limit: 1000
    });

    // Count messages per collaborator
    const counts = new Map<string, number>();

    collaborations.forEach(m => {
      const partner = m.metadata?.toMemorySpace as string;
      if (partner) {
        counts.set(partner, (counts.get(partner) || 0) + 1);
      }
    });

    // Add edges
    for (const [partner, count] of counts.entries()) {
      network.edges.push({
        from: space.memorySpaceId,
        to: partner,
        weight: count,
        type: 'COLLABORATED_WITH'
      });
    }
  }

  return network;
}

// Result: Memory space collaboration graph
{
  nodes: [
    { id: 'finance-space', name: 'Finance Agent', type: 'custom' },
    { id: 'hr-space', name: 'HR Agent', type: 'custom' },
    { id: 'legal-space', name: 'Legal Agent', type: 'custom' }
  ],
  edges: [
    { from: 'finance-space', to: 'hr-space', weight: 45 },
    { from: 'finance-space', to: 'legal-space', weight: 23 },
    { from: 'hr-space', to: 'legal-space', weight: 12 }
  ]
}

Visualization:

finance-agent ══(45)══> hr-agent
      ║                    │
     (23)                 (12)
      ║                    │
      ╚══════════> legal-agent

Use Cases:

  • Find bottleneck agents (highest edge weight)
  • Identify collaboration patterns
  • Recommend agent for task based on network position

Performance Characteristics​

Graph-Lite Performance​

HopsQuery TypeAvg LatencyComplexityUse Case
1Single lookup10-50msSimpleDirect relationships
2-3Sequential queries50-200msModerateContext chains, A2A patterns
4-5Multiple queries200-500msComplexWorkflow hierarchies
6+Many queries500-2000ms+Very complexConsider native graph

Example:

// 3-hop query: User → Conversation → Context → Agent
// Query 1: Get conversation (by userId)
const convos = await cortex.conversations.list({ userId }); // 20ms

// Query 2: Get contexts triggered by conversation
const contexts = await cortex.contexts.search({
  "conversationRef.conversationId": convos[0].conversationId,
}); // 30ms

// Query 3: Get agent handling context
const memorySpace = await cortex.memorySpaces.get(contexts[0].memorySpaceId); // 15ms

// Total: ~65ms (acceptable)

Scaling:

  • Well-indexed queries remain fast even with millions of documents
  • Agent isolation helps (queries scoped to single agent are fastest)
  • Compound indexes critical (e.g., by_agent_userId)

When Graph-Lite Is Sufficient​

Use Cases That Work Well:

  1. Context Chain Hierarchies (depth typically 1-5)

    • Task delegation workflows
    • Approval chains
    • Project/sprint structures

  2. Direct Relationships (1-2 hops)

    • "Get user's conversations"
    • "Find agent's active contexts"
    • "List memories for this context"

  3. Audit Trails (3-5 hops)

    • Conversation → Context → Memories → Agent
    • Compliance queries with known path

  4. Agent Collaboration (1-3 hops)

    • Direct collaborators
    • Recent communication patterns
    • Message threads

Performance Profile:

  • Queries complete in <500ms
  • Acceptable for user-facing features
  • Good enough for background jobs
  • Pagination helps with large result sets

When to Add a Graph Database​

Limitations of Graph-Lite:

  1. Deep Traversals (6+ hops)

    • Each hop = another Convex query
    • Latency compounds (500ms+ becomes unusable)
    • Memory consumption increases

  2. Complex Patterns

    • "Find all paths between A and B"
    • "Which nodes are most central?" (PageRank)
    • Pattern matching with wildcards

  3. Graph Algorithms

    • Shortest path algorithms
    • Community detection
    • Centrality calculations
    • Recommendation graphs

  4. Highly Connected Graphs

    • Social networks (users connected to users)
    • Knowledge graphs with dense relationships
    • Recommendation engines

Use Cases Requiring Native Graph:

  • Healthcare: Clinical trial relationships, drug interactions, patient histories
  • Legal: Case law precedents, evidence chains, multi-party litigation
  • Finance: Transaction networks, fraud detection, risk graphs
  • Research: Citation networks, collaboration graphs, concept relationships

Decision Matrix:

Your NeedGraph-LiteNative Graph DB
Context hierarchies (depth <5)PerfectOverkill
Audit trails (known paths)PerfectOverkill
Agent collaboration (1-3 hops)GoodBetter
Deep traversals (6+ hops)Too slowRequired
Pattern matchingVery hardRequired
Graph algorithmsNot feasibleRequired
Dense relationship networksSlowBest

When to Upgrade:

  • Queries consistently taking >500ms
  • Need for complex pattern matching
  • Graph algorithm requirements
  • Dense relationship graphs (>1M edges)

Transition to Native Graph​

Step 1: Identify Need​

Monitor query performance:

// Track graph-like query latency
async function monitoredTraversal(startId: string, depth: number) {
  const start = Date.now();

  const result = await traverseContextChain(startId, depth);

  const latency = Date.now() - start;

  if (latency > 500) {
    console.warn(`Graph-Lite query slow: ${latency}ms at depth ${depth}`);
    console.warn("Consider upgrading to native graph database");
  }

  return result;
}

Step 2: Evaluate Options​

See: Graph Database Integration Guide for complete setup and integration instructions.

Quick comparison:

DatabaseSetup DifficultyPerformanceCostBest For
Neo4j CommunityMedium (Docker)ExcellentFree (self-hosted)Production ready
MemgraphEasy (Docker)ExcellentFree (self-hosted)High performance

Step 3: Migration Path​

Parallel Operation:

// Run both Graph-Lite and native graph during migration
const cortex = new Cortex({
  convexUrl: process.env.CONVEX_URL,

  graph: {
    enabled: true,
    provider: 'neo4j',
    connection: { uri: 'bolt://localhost:7687', ... },

    // Compare results during migration
    compareMode: true,  // Query both, compare latency
    fallbackToConvex: true  // Use Graph-Lite if graph fails
  }
});

Verify Results Match:

// Ensure graph and Convex give same results
const graphResult = await cortex.graph.traverse(...);
const convexResult = await cortexTraverse(...);

assert.deepEqual(graphResult, convexResult);

Cutover:

// Once confident, switch fully to graph for complex queries
const cortex = new Cortex({
  graph: {
    enabled: true,
    preferGraphFor: ["traverse", "findPath", "pattern"],
    fallbackToConvex: true, // Safety net
  },
});

Advanced Graph-Lite Patterns​

Pattern: Time-Based Graph Queries​

// Find what was connected at a specific point in time
async function getHistoricalGraph(timestamp: Date) {
  // Get contexts that existed at that time
  const contexts = await cortex.contexts.list({
    createdBefore: timestamp,
    $or: [
      { completedAfter: timestamp },
      { status: { $in: ["active", "blocked"] } },
    ],
  });

  // Get A2A messages from that period
  const a2aMessages = await cortex.memory.search("*", "*", {
    source: { type: "a2a" },
    createdBefore: timestamp,
    createdAfter: new Date(timestamp.getTime() - 30 * 24 * 60 * 60 * 1000), // 30 days window
  });

  // Reconstruct graph state at that time
  return {
    activeContexts: contexts,
    activeCollaborations: buildCollaborationGraph(a2aMessages),
    timestamp,
  };
}

Pattern: Weighted Graph Edges​

// Build graph with weighted edges (importance, frequency, recency)
async function getWeightedCollaborationGraph() {
  const spaces = await cortex.memorySpaces.list();
  const edges = [];

  for (const space of spaces.spaces) {
    const a2a = await cortex.vector.search(space.memorySpaceId, "*", {
      sourceType: "a2a",
      limit: 1000,
    });

    // Calculate edge weights
    const weights = new Map();

    a2a.forEach((m) => {
      const partner = m.metadata?.toMemorySpace as string;
      if (!partner) return;

      const current = weights.get(partner) || {
        count: 0,
        totalImportance: 0,
        lastContact: 0,
      };

      weights.set(partner, {
        count: current.count + 1,
        totalImportance: current.totalImportance + (m.importance || 0),
        lastContact: Math.max(current.lastContact, m.createdAt),
      });
    });

    // Create weighted edges
    for (const [partner, stats] of weights.entries()) {
      const recencyWeight = Math.max(
        0,
        1 - (Date.now() - stats.lastContact) / (30 * 24 * 60 * 60 * 1000),
      );
      const importanceWeight = stats.totalImportance / stats.count / 100;
      const frequencyWeight = Math.min(stats.count / 100, 1);

      edges.push({
        from: space.memorySpaceId,
        to: partner,
        weight: (recencyWeight + importanceWeight + frequencyWeight) / 3,
        count: stats.count,
        avgImportance: stats.totalImportance / stats.count,
        lastContact: new Date(stats.lastContact),
      });
    }
  }

  return edges.sort((a, b) => b.weight - a.weight);
}

Best Practices​

1. Use Existing Traversal Methods​

// Good: Use built-in traversal when available
const chain = await cortex.contexts.get(contextId, { includeChain: true });

// Avoid: Manual traversal when built-in exists
const context = await cortex.contexts.get(contextId);
const parent = context.parentId
  ? await cortex.contexts.get(context.parentId)
  : null;
const grandparent = parent?.parentId
  ? await cortex.contexts.get(parent.parentId)
  : null;
// etc. (slower and more code)

2. Limit Traversal Depth​

// Good: Set reasonable depth limits
const MAX_DEPTH = 5;

async function safeTraverse(contextId: string, depth = 0) {
  if (depth > MAX_DEPTH) {
    console.warn("Max depth reached, consider graph database");
    return [];
  }

  // ... traverse logic
}

3. Cache Traversal Results​

// Cache frequently-accessed graphs
const graphCache = new Map();

async function cachedContextChain(contextId: string) {
  if (graphCache.has(contextId)) {
    return graphCache.get(contextId);
  }

  const chain = await cortex.contexts.get(contextId, { includeChain: true });

  graphCache.set(contextId, chain);
  setTimeout(() => graphCache.delete(contextId), 60000); // 1 min TTL

  return chain;
}

4. Use Batch Queries​

// Good: Fetch all related entities in one query when possible
const contexts = await cortex.contexts.list({
  rootId: rootContextId, // Gets entire workflow in one query
});

// Avoid: Querying one-by-one
const root = await cortex.contexts.get(rootId);
for (const childId of root.childIds) {
  const child = await cortex.contexts.get(childId); // N queries!
}

Comparison: Graph-Lite vs Native Graph​

FeatureGraph-Lite (Built-in)Native Graph DB
SetupNone (included)Docker or managed service
StorageConvex onlyConvex + Graph DB
ConsistencyAlways consistentEventually consistent
Query Speed50-500ms (depth-dependent)10-100ms (constant)
Max Depth5 hops practical100+ hops
AlgorithmsManual implementationBuilt-in
CostFree (Convex storage)Graph DB cost + sync
ComplexityLow (no extra systems)Medium (2 databases)
Use Cases90% of use casesAdvanced/enterprise

Summary: Start with Graph-Lite. Upgrade to native graph only when you hit performance/complexity limits.

Migration Example​

Phase 1: Using Graph-Lite

// Current code works fine
const chain = await cortex.contexts.get(contextId, { includeChain: true });
// 150ms, depth 3 - acceptable

Phase 2: Hit Limits

// Performance degrading
const chain = await cortex.contexts.get(complexContextId, {
  includeChain: true,
});
// 800ms, depth 8 - too slow

// Complex query needed
// "Find all agents who worked on contexts related to user-123 through any path"
// → Very difficult with Graph-Lite, easy with native graph

Phase 3: Add Native Graph

// Enable graph integration
const cortex = new Cortex({
  convexUrl: process.env.CONVEX_URL,
  graph: {
    enabled: true,
    provider: 'neo4j',
    connection: { uri: 'bolt://localhost:7687', ... }
  }
});

// Now use graph for complex queries
const agents = await cortex.graph.traverse({
  start: { type: 'user', id: 'user-123' },
  relationships: ['INVOLVES', 'HANDLED_BY'],
  maxDepth: 10
});
// 45ms - fast!

// Keep using Graph-Lite for simple queries
const chain = await cortex.contexts.get(contextId, { includeChain: true });
// Still works, automatically uses best method

Next Steps​

Questions? Ask in GitHub Discussions.