The context a support agent actually needs
None of these are documents. They are state about a relationship, and they change. That is why a document search alone cannot keep a support agent oriented.
- Account facts: plan, seats, billing interval and renewal date.
- Open issues: the ticket, the workaround already tried and the current status.
- Preferences: channel, time zone and the level of detail the customer wants.
- Escalation rules: who signs off, who is on call and what must never be promised.
- Commitments: what the team said it would do, and by when.
What to store, and what to redact
| Store | Redact or avoid |
|---|---|
| Plan, seats and renewal date. | Full card numbers or bank details. |
| Open ticket and failed workaround. | Passwords, tokens and private keys. |
| Escalation owner and approval limit. | Unrelated health or family details. |
| Preferred channel and contact time. | Raw transcripts that contain sensitive data. |
| Commitments and deadlines. | Speculation about the customer’s intent. |
Scopes that keep one customer out of another
Filter at read time, not in the prompt. A prompt instruction is not an access control. If the retrieval query does not carry the tenant and user scope, the wrong memory can appear in the context.
| Scope | Example | Why it matters |
|---|---|---|
| Tenant | tenant:northwind | Account facts and isolation between companies. |
| User | user:dana | Personal preferences and contact rules. |
| Agent | agent:support-bot | What the support bot learned while working the ticket. |
| Session | session:ticket-4821 | Temporary state for the current conversation. |
A support workflow with memory
- Before the reply: retrieve account facts, open issues, preferences and the last commitment.
- During the reply: use the current ticket and the retrieved memories, not the entire account history.
- After resolution: write the outcome, the workaround that worked and any new commitment.
- When a fact changes: supersede the old memory and keep the version for audit.
- When a customer leaves: delete by user or tenant and verify that derived memories are gone.
context = memory.search(
"what should we know before replying to this customer?",
user_id="dana",
agent_id="support-bot",
metadata={"tenant": "northwind"},
limit=6,
)Metrics that show whether memory is helping
| Metric | What it tells you |
|---|---|
| Repeat-question rate | How often the agent asks for information it already has. |
| First-contact resolution | Whether the agent has the context to solve the issue sooner. |
| Escalation accuracy | Whether the right cases reach the right owner. |
| Contradiction rate | Whether the agent repeats a superseded plan or promise. |
| Handle time | Whether the agent spends less time reconstructing context. |
| Deletion correctness | Whether departed customers disappear from retrieval. |
Privacy is part of the support experience
Support data is some of the most sensitive data in a company. A memory layer should make retention, redaction and deletion easier, not harder.
- Set a TTL for session memories and short-lived ticket state.
- Delete by user or tenant when an account closes.
- Keep an audit trail that proves deletion without keeping the content.
- Redact sensitive fields before the write.
- Export the graph when a customer requests their data.
Frequently asked questions
Can AI Memory replace our help desk?
No. It is a memory layer beside your help desk and support agent. The ticket system remains the system of record for the conversation and the workflow.
What should a support agent remember first?
Start with account facts, open issues, failed workarounds, escalation rules and commitments. Those are the details that most often force a customer to repeat themselves.
How do you prevent cross-customer leakage?
Carry tenant and user scope on every write and filter on them at read time. Do not rely on prompt instructions for isolation.
How long should support memory be retained?
It depends on your policy and the customer relationship. Use TTLs for session state, a defined retention window for ticket memories and deletion by user or tenant when an account closes.
Can we store sensitive information in memory?
Store only what is necessary and redact the rest before the write. A memory layer persists data, so it inherits your retention, backup and access obligations.
How do we know memory improves support?
Measure repeat questions, first-contact resolution, escalation accuracy, contradiction rate and handle time on a small set of real conversations before and after memory.