The Conversation Graph The system of record for conversations.

One thread per customer. Every fact with its source. Shared by every agent, every person and every system that works with that customer.

Book a technical deep dive Read the architecture

John Doe Sample customer

Customer since: 12 Mar
City: Pune
Language: Hinglish
Prefers: WhatsApp, evenings

01 / 14moments on the thread

Held by 4 AI · 1 person

One thread · 12 Mar to today, then what's plannedClick a card

Today

AdWhatsAppVoiceEmailYour systems

Built around the conversation, not the record.

Most systems put the record first and attach the conversation to it. We invert that, because the conversation is where intent, decisions and context originate.

Record-first

CallNoteContact record

History becomes attachments. Context is lost at every handoff.

Conversation-first

ConversationFactsYour records

Records are updated from what was said.

And no, agent memory doesn't fix this. It's private to one agent. Memory remembers. The graph records.

Why now

Customers message instead of clicking.
Agents now do work that spans weeks.
Connecting systems is configuration, not a project.

Five principles.

01
One person, one thread
However many channels, agents and teams they touch.
02
Every fact has a source
The turn, the channel, the time and who said it.
03
Conversations are data, not logs
Structured as they happen, not mined later.
04
Context is assembled, not dumped
Each agent gets the slice it needs, within its permissions.
05
Write back, never lock in
Outcomes flow to your systems; your data stays exportable.

A context graph, with the conversation as its primary object.

Most context graphs are built from documents and entities. This one is built from what customers, agents and people actually said and did.

Ingestchannels · systems · actions

Resolveone person, every identity

Extractturns → facts, with source

Store · The graph

PersonAgent · Rep

↓ participate in

Conversationprimary object

↓ contains

TurnFactYour objects

Serve

Agents at runtime
Queries · Zig AI
Real-time triggers
Sync to your systems
Warehouse export

Governance · every layerProvenanceActor attributionAccess scopesPII redactionRetentionResidency

One turn, end to end.

01
Message arrives on any channel.
02
Person resolved to their one thread.
03
Context assembled: the slice this agent needs and may see.
04
Agent acts: replies, takes actions in your systems.
05
New facts written with source and actor.
06
Systems synced: CRM and objects updated.

Sample context packet · step 3 · illustrative

{
  "conversation_id": "cv_[id]",
  "person": {
    "id": "p_[id]",
    "channels": ["whatsapp", "voice", "email"],
    "preferred_channel": "whatsapp"
  },
  "relationship": {
    "agent": "onboarding",
    "stage": "documents_pending"
  },
  "facts": [
    {
      "key": "[field]",
      "value": "[value]",
      "source": "turn t_[n] · whatsapp",
      "said_by": "person",
      "at": "[timestamp]",
      "confidence": 0.94
    }
  ],
  "objects": { "[YourObject]": { "status": "[value]" } },
  "open_items": ["[pending step]"]
}

Where it sits next to memory, context graphs, CRMs and CDPs.

	Agent memory	Vector store	Context graph	CRM	CDP	Conversation Graph
Primary object	Agent session	Embeddings	Documents, entities	Record	Profile, events	Conversation
Shared across agents and people	No	Partial	Yes	People only	Marketing only	Yes
Full conversation turns	Partial	Chunks	Rarely	Notes	No	Yes
Every fact traceable to source	No	No	Partial	No	Partial	Yes
Structured, queryable facts	No	Similarity only	Yes	Fields only	Events only	Yes
Real-time read and write by agents	One agent	Read	Varies	Via API	Webhooks	Native
Records who did what (human, agent, system)	No	No	Varies	Humans only	No	Yes

Describes typical implementations.

It doesn't replace your CRM or CDP. It captures the layer they never held, and writes the results back into them.

Agents run on it. Teams ask it questions.

Ask in plain language

Zig AI answers questions across every conversation.

Who raised pricing concerns this week?

Trigger on what was said

Start a workflow when intent, a fact or a stage changes.

Attribute end to end

From ad click through every conversation to revenue.

Export to your warehouse

Structured conversation data for your BI and models.

Beyond the customer journey

The same graph works for any process run through conversation.

Field and distributed sales teams
Partner and dealer networks
Citizen services
Internal and employee support

Accountability is what lets agents do real work.

When AI does real work, every change needs an owner. The graph records who did what, whether a person, an agent or an external system.

ProvenanceEvery fact linked to the turn it came from.

Actor attributionEvery action tagged human, agent or system.

Access scopesEach agent and team sees only what it should.

PII handling and retentionRedaction and retention controls.

Residency and deploymentOur cloud in your region, your cloud, or on-prem.

Never trained on your dataYour conversations stay yours.

Open by design.

APIs

Read and write conversations, facts and objects.

Webhooks

Subscribe to facts, stage changes and actions.

Schema

Your data model as JSON, versioned.

Warehouse export