Customer Memory Graph: How Aartha Builds It

Most customer tools are filing cabinets. They store what happened — a meeting here, an email there, a CRM field somewhere else — and leave it to you to reconstruct what it all means before every call. The Customer Memory Graph flips that. Instead of storing records, it remembers facts: who the buyer is, what's at risk, what was committed, and — crucially — what changed.

What is a Customer Memory Graph?

A Customer Memory Graph (CMG) is a bi-temporal, cited graph of facts about each account. Three ideas make it different from a CRM or a dashboard:

1. Facts, not rows. "Maria Chen is the economic buyer at Acme" is a fact — with a confidence score, a date it became true, and a source. It isn't a field you overwrite.
2. Bi-temporal history. When a later meeting reveals a new buyer, Aartha doesn't delete the old fact. It closes it and records the change. That history is the product: you can ask "who was the buyer, and when did that change?"
3. Provenance on everything. Every fact links back to the exact meeting utterance or email it came from. Answers come with receipts.

Why this beats a dashboard

Dashboards answer "what is my book worth?" They can't answer "what needs me today and why?" — because that question is about change, and a snapshot has no memory of yesterday.

Because the CMG records change as a first-class event, Aartha can surface a real diff of your book: a champion who changed, a commitment that slipped, sentiment that reversed. Each one is cited, so you can trust it without re-reading the thread.

How Aartha builds it — automatically

You don't maintain the graph. Aartha does, off the signals you already generate:

• Extract. After each meeting or email is analyzed, an LLM proposes candidate facts (subject, predicate, object) with a supporting quote.
• Resolve. Each entity is matched to the existing graph by meaning, not just string equality, so "Maria" and "Maria Chen" collapse to one node.
• Reconcile. New facts are compared to what's already known. Aartha decides: add, corroborate, or invalidate-and-replace. Contradictions become citable "what changed" events instead of silent overwrites.
• Retrieve. When you ask a question or open an account, Aartha pulls the most relevant facts — vector + keyword + graph hops — and answers with sources.

What it unlocks

Once the memory exists, every other surface gets sharper:

• Risk detection diffs facts, so you get "the champion changed" instead of a vague "negative email."
• The assistant answers grounded and cited, and remembers across sessions.
• Drafts — renewal emails, CRM updates — are grounded in real, cited context, then gated by your approval.

The result is a platform that doesn't just store your customer data. It actually knows your customers — and tells you the moment something changes.