RAG vs Memory for a Support Agent: Which Should You Use?
Guide · · Published November 2026 · 5 min read
By Ashwin Gopinath, CEO
TL;DR
For a support agent, use both. RAG answers from documents every customer shares, such as help center articles and policies. Memory answers what is true about this customer now: what was promised, what was tried, which exception applies. Sentra is the memory side, built from Slack, email, meetings, the CRM and project tools and shared by agents and people.
What each one is for
RAG retrieves text that looks similar to the question at the moment it is asked. That suits "how do I reset two-factor authentication", where the answer is in a help article and is the same for every customer. Keep the article current and RAG stays correct.
Memory decides what is true when information arrives, remembers when it changed, and serves the current fact on every later call. That suits "why was this customer told they could have a refund outside the window", where the answer is an exception agreed on a call in March and nowhere in the help center.
Which to use, by the question the agent gets
| The customer asks | Where the answer lives | Use |
|---|---|---|
| How does a feature work? | Help center and product docs | RAG |
| What is your refund policy? | A published policy page | RAG, with memory if the policy changed recently |
| What did your colleague promise me last week? | A call, an email or a ticket note | Memory |
| Is the fix you mentioned shipped yet? | Engineering's tracker and chat | Memory |
| Why is my contract different from the standard one? | The sales conversation and the CRM | Memory |
| Who is handling my account now? | The CRM, as of today | Memory |
The pattern is simple. If the answer is the same for every customer and lives in a document you control, retrieve it. If the answer depends on this customer's history or on something that changed, the agent needs memory.
What goes wrong with RAG alone
A support agent that answers a ticket correctly on Monday and contradicts itself on the same account by Friday does not have a knowledge problem. It has a memory problem. Three failures show up first in support. The agent quotes a policy that changed, because the old and new pages both match the question. It asks the customer to repeat what they already told a colleague, because the earlier conversation is in another system. And it misses the exception, because exceptions are agreed in conversations and never written into the knowledge base. The full list of failure modes is in AI agent memory vs RAG.
What Sentra provides as the memory side
Sentra is the managed organization memory for teams and AI agents. For support and success teams it builds a shared memory of every account from Slack, email, meetings, the CRM and project tools: what was promised, what shipped, what the customer said on calls, which policies changed and who owns what. Each fact is stored once with its source and the date it changed.
When a newer statement contradicts an older one, the older fact gets an end date and stays as history, so the agent quotes the current policy and can still explain what applied when the customer signed. A support agent reads the memory over MCP or REST alongside its help center retriever. Answers respect the permissions of the system each fact came from, and a failed check counts as no access. The human team asks the same memory in Slack, so the agent and the person who picks up the escalation are working from one account history.
It complements a help center or ticketing knowledge base and does not replace one. The help center answers customers. The memory answers the team and its agents about each customer. Our guide to a company brain for customer support covers the rollout.
When something else fits better
If your support agent only answers how-to questions from public documentation, RAG is all you need. If you are building a consumer chatbot that should remember each end user's preferences inside your own app, a per-user memory API such as Mem0 or Zep is designed for that and is written to directly from your code. Sentra's memory is built from your company's connected systems and its public API is read-oriented, so it is the right fit for what your company knows about a customer, and not for a chatbot's own notes about a conversation. If your ticketing platform's built-in AI covers your cases, start there; our note on Zendesk AI explains what it sees.
How to combine them
1. Keep RAG over the help center and policies, and make sure retired articles are removed from the index.
2. Add memory as a second tool the agent calls for anything about a named customer, a commitment or a recent change.
3. Let memory win when they disagree. If the policy page says 30 days and a dated decision says 14 from this month, the agent should use the decision and cite it.
4. Test with a real escalation. Take a ticket that needed three people to resolve and see whether the agent now has what they knew.
To try that on your own accounts, book a demo.
FAQ
RAG vs memory: which should I use for a support agent?
Both. Use RAG for help center articles and published policies that are the same for every customer. Use memory, such as Sentra, for what was promised to a specific customer, what changed and who owns the account.
Can memory replace the knowledge base for support?
No. Product documentation belongs in a help center, and retrieval over it works well. Memory adds the account history and recent decisions the help center never contains.
How does a support agent avoid quoting an outdated policy?
By reading policy changes from a memory that dates each fact. Sentra end-dates the older policy when a newer decision replaces it, so the agent gets the current one and can still say what applied before.
Does the support agent need to be rebuilt to add memory?
No. Sentra is added as one more tool over MCP or REST next to the existing retriever.