Conversation History

Layer 1a: Conversations

Conversation History is the ACID source of truth for all message data. Every message is immutably stored—append-only, never modified or deleted—providing complete audit trails.

---

Quick Start

// Create a conversation
const conv = await cortex.conversations.create({
  type: 'user-agent',
  memorySpaceId: 'user-123-space',
  participants: {
    user: { id: 'user-123', name: 'Alice' },
    agent: { id: 'assistant' },
  },
});

// Add messages (immutable, append-only)
await cortex.conversations.addMessage({
  conversationId: conv.id,
  message: {
    role: 'user',
    content: 'Hello!',
    userId: 'user-123',
  },
});

await cortex.conversations.addMessage({
  conversationId: conv.id,
  message: {
    role: 'agent',
    content: 'Hi there! How can I help?',
  },
});

// Retrieve history
const history = await cortex.conversations.getHistory(conv.id, {
  limit: 50,
  order: 'desc',
});
// history.messages contains the array of messages

---

Where Conversations Fit

Conversations live in Layer 1a of Cortex's 4-layer architecture. They work together with Layer 2 (Vector) for efficient retrieval:

Feature	Layer 1a: ACID Conversations	Layer 2: Vector Index
Purpose	Source of truth	Searchable index
Mutability	Immutable (append-only)	Versioned with updates
Retention	Forever (compliance)	Configurable limits
Search	Chronological retrieval	Semantic search
Use Case	Audit trails, replay	Fast context building

The Link Between Layers

Vector memories reference ACID conversations via conversationRef. You can always trace back from any memory to its original conversation.

---

Conversation Operations

// User-Agent conversation
const conv = await cortex.conversations.create({
  type: 'user-agent',
  memorySpaceId: 'user-123-space',
  participants: {
    user: { id: 'user-123', name: 'Alice' },
    agent: { id: 'assistant' },
  },
  metadata: { channel: 'web' },
});

// Agent-Agent conversation (A2A) - manual approach
// Note: cortex.a2a.send() handles this automatically (recommended)
const a2aConv = await cortex.conversations.create({
  type: 'agent-agent',
  participants: {
    agent1: 'finance-agent',
    agent2: 'hr-agent',
  },
});

// User message
const msg = await cortex.conversations.addMessage({
  conversationId: conv.id,
  message: {
    role: 'user',
    content: 'What is the weather?',
    userId: 'user-123',
  },
});

// Agent response
await cortex.conversations.addMessage({
  conversationId: conv.id,
  message: {
    role: 'agent',
    content: 'The weather is sunny and 72°F.',
  },
});

// Messages are NEVER modified - append-only
console.log(msg.id);  // Use in conversationRef

// Get recent messages
const history = await cortex.conversations.getHistory(conv.id, {
  limit: 50,
  order: 'desc',  // Most recent first
});

history.messages.forEach(msg => {
  console.log(`${msg.role}: ${msg.content}`);
});

// Get full conversation
const full = await cortex.conversations.get(conv.id);
console.log(full.messageCount);

// List user's conversations
const convs = await cortex.conversations.list({
  type: 'user-agent',
  'participants.userId': 'user-123',
  sortBy: 'lastMessageAt',
  sortOrder: 'desc',
});

// Search across conversations
const results = await cortex.conversations.search(
  'user-123',
  'pricing discussion',
  { limit: 10 }
);

---

Message Schema

Parameter	Type	Required	Default	Description
role	string	Yes	—	'user' | 'agent' | 'system'
content	string	Yes	—	Message content
userId	string	No	—	User who sent (for user messages)
timestamp	Date	No	—	Auto-set if not provided
metadata	object	No	—	Custom message data (channel, device, etc.)

---

Conversation Types

Standard chat between user and AI:

{
  conversationId: 'conv-456',
  type: 'user-agent',
  participants: {
    userId: 'user-123',
    memorySpaceId: 'agent-1'
  },
  messages: [
    { id: 'msg-001', role: 'user', content: '...', timestamp: T1 },
    { id: 'msg-002', role: 'agent', content: '...', timestamp: T2 }
  ]
}

Communication between agents. Recommended: Use cortex.a2a.send() which handles conversation creation automatically:

// Recommended: A2A helper handles ACID conversation automatically
await cortex.a2a.send({
  from: 'finance-agent',
  to: 'hr-agent',
  message: 'What is the Q4 budget?',
  importance: 85,
});
// Creates ACID conversation + stores in both agents' memories

The resulting ACID conversation structure:

{
  conversationId: 'a2a-conv-789',
  type: 'agent-agent',
  participants: {
    agent1: 'finance-agent',
    agent2: 'hr-agent'
  },
  messages: [
    { id: 'a2a-msg-001', from: 'finance-agent', to: 'hr-agent', content: '...' }
  ]
}

See A2A Communication for full API reference.

---

Immutable Audit Trail

ACID Properties

• Atomicity: Message fully stored or not at all
• Consistency: Conversation always in valid state
• Isolation: Concurrent messages don't interfere
• Durability: Once stored, never lost

Compliance Benefits

Messages are never modified or deleted from Layer 1a. Vector memories (Layer 2) can be deleted while preserving the ACID conversation history for compliance.

---

Streaming Support

For Streaming LLM Responses

Use rememberStream() which handles progressive storage automatically:

import { streamText } from 'ai';
import { openai } from '@ai-sdk/openai';

const response = await streamText({
  model: openai('gpt-4'),
  messages: [{ role: 'user', content: 'What is AI?' }],
});

// Automatically buffers and stores
const result = await cortex.memory.rememberStream({
  memorySpaceId: 'my-agent',
  conversationId: conv.id,
  userMessage: 'What is AI?',
  responseStream: response.textStream,
  userId: 'user-123',
  userName: 'Alice',
});

console.log(result.fullResponse);

See Streaming Support for full documentation.

---

Building Context

Recent Context (Chronological)

async function buildRecentContext(conversationId: string) {
  const history = await cortex.conversations.getHistory(conversationId, {
    limit: 10,
  });

  return history.messages
    .map(msg => `${msg.role === 'user' ? 'User' : 'Assistant'}: ${msg.content}`)
    .join('\n');
}

Relevant Context (Semantic)

async function buildRelevantContext(memorySpaceId: string, query: string) {
  // Search vector memories
  const memories = await cortex.memory.search(memorySpaceId, query, {
    embedding: await embed(query),
    limit: 5,
  });

  // Retrieve full conversations via conversationRef
  for (const memory of memories) {
    if (memory.conversationRef) {
      const conv = await cortex.conversations.get(
        memory.conversationRef.conversationId
      );
      // Access full context...
    }
  }
}

---

Common Patterns

Multi-Session Continuity

async function handleReturningUser(userId: string, message: string) {
  // Find recent conversations
  const convs = await cortex.conversations.list({
    'participants.userId': userId,
    sortBy: 'lastMessageAt',
    sortOrder: 'desc',
    limit: 1,
  });

  const lastConv = convs[0];
  const timeSinceLast = lastConv
    ? Date.now() - lastConv.lastMessageAt.getTime()
    : Infinity;

  // Continue or start new
  const conversationId = timeSinceLast < 30 * 60 * 1000
    ? lastConv.id
    : (await cortex.conversations.create({
        type: 'user-agent',
        participants: { userId, agentId: 'assistant' },
      })).id;

  await cortex.conversations.addMessage({
    conversationId,
    message: {
      role: 'user',
      content: message,
      userId,
    },
  });

  return conversationId;
}

Context Window Management

async function buildContextWindow(conversationId: string, maxTokens: number) {
  const history = await cortex.conversations.getHistory(conversationId);
  const messages = history.messages;
  
  const contextMessages = [];
  let tokenCount = 0;

  // Build from newest to oldest
  for (let i = messages.length - 1; i >= 0; i--) {
    const tokens = estimateTokens(messages[i].content);
    if (tokenCount + tokens > maxTokens) break;
    contextMessages.unshift(messages[i]);
    tokenCount += tokens;
  }

  return contextMessages;
}

---

Best Practices

Store Complete Metadata

await cortex.conversations.addMessage({
  conversationId: conv.id,
  message: {
    role: 'user',
    content: message,
    userId,
    metadata: {
      channel: 'web',
      device: 'desktop',
      sessionId: 'session-123',
    },
  },
});

Redact Before Storing

// ACID is immutable - redact BEFORE storing
function redact(text: string) {
  return text
    .replace(/\b\d{3}-\d{2}-\d{4}\b/g, '[SSN REDACTED]')
    .replace(/\b\d{16}\b/g, '[CARD REDACTED]');
}

await cortex.conversations.addMessage({
  conversationId: conv.id,
  message: {
    role: 'user',
    content: redact(userMessage),
    userId,
  },
});

</Callout>

<Callout type="tip" title="Summarize Long Conversations">

```typescript
// For long conversations, store summaries as vector memories
if (conversation.messageCount > 100) {
  const summary = await summarize(olderMessages);
  
  await cortex.memory.remember({
    memorySpaceId,
    content: `Summary: ${summary}`,
    conversationRef: { conversationId, messageIds: [] },
  });
}

---

Next Steps