Memory for support agents

Customer context is not blanket permission.

A support agent may remember why a customer contacted you. Before it issues a credit, changes a plan, or closes a case, it should prove which memory supports that action.

The gap

A useful inference can still be unsafe to act on.

“The customer sounds frustrated” can help shape a reply. It should not be treated like “the customer asked to cancel.” ContextDB keeps source and confidence attached so policy can tell the difference.

# good context, not action authority
{
  "content": "Customer may want to cancel",
  "source": "agent_inferred",
  "confidence": 0.62
}
→ useful for conversation, held from action

# direct customer instruction
{
  "content": "Cancel at the end of this cycle",
  "source": "user_stated",
  "confidence": 0.98
}
→ evaluate against cancellation policy

Consequential support actions

Use action recall where a mistake changes the account.

Billing

Credits and refunds

Verify the reason and user-stated instruction before a support agent changes a financial record.

action → issue_credit

Account

Plans and permissions

Distinguish a request, a historical preference, and an agent inference before mutating account state.

action → change_plan

Case

Resolution and closure

Keep the evidence that justified an automated resolution so operators can review what happened later.

decision → reason + evidence IDs

Keep your system of record

ContextDB does not own the customer account.

Your CRM, billing system, help desk, or database remains authoritative for account state. ContextDB stores the remembered evidence and the policy decision that your support host uses before a mutation.

See the architecture →

help desk / CRM / billing
customer and account records
             ▲
             │ host mutation
             │
support agent host
             ▲
             │ act / ask / abstain
             │
ContextDB
memories + evidence + decision record

Questions

What support teams should know.

Does ContextDB decide our refund policy?

No. Your team defines business rules and host enforcement. ContextDB evaluates remembered evidence under the active trust policy.

Do we keep our help desk and CRM?

No. It complements those systems by tracking what an agent remembers and why that memory was or was not eligible to drive an action.

What happens when evidence is missing?

The safe result is ask or abstain. The host should request a specific confirmation or route the case to a person.

Start with one support action that needs evidence.

Model the source, run action recall, and make the host respect the result.

Try it locally