WhatsApp AI assistant memory audit checklist: what it remembers, what it forgets, and how to verify it

8 min read

When an AI assistant handles WhatsApp, memory is what decides whether the customer feels known or feels like they're being asked the same thing three times. The problem is that "memory" isn't a single thing: there's context within a conversation, data stored on the contact, and something the assistant learned from previous replies. If you don't separate those layers, you can't audit anything. This checklist organizes the tests by stage and tells you, for each one, how to know it's working.

The audit is done on observable behavior: what the assistant replies, what gets written to the contact record, and what happens when the person handling the conversation changes shifts. No access to code or internal configuration is needed. If something can't be seen from the outside, it isn't auditable and doesn't work as a test.

Before you start: separate the three layers of memory

Most audits fail because they mix up three different things and then don't know what to fix. Before running any test, make it clear which one you're measuring.

Session memory: what the assistant remembers within a single conversation. If the customer said "I'm looking for an appointment on Thursday" and three messages later the assistant asks "what day works for you?", session memory failed.
Contact memory: what gets stored on the customer record and survives the conversation closing. Name, tags, last order, stated preferences.
Learned memory: what the assistant picked up from the team's real replies and now suggests differently. It isn't audited by reading a conversation: it's audited by comparing suggestions for the same situation at two different points in time.

If your assistant doesn't distinguish these layers, the audit still works: the result will be that everything behaves like session memory and nothing persists. That in itself is a finding.

Stage 1: session memory audit

It's tested with a single conversation, without closing it. The goal is to see whether the assistant keeps the thread when the customer gives information in installments, changes topic and comes back, or writes in several messages in a row.

Run each test in a new conversation so the result isn't contaminated by previous history.

Test	How it's run	Sign that it's OK
Data split across messages	The customer gives the data spread across three messages (for example, product, quantity and area in separate messages).	The assistant builds the complete request without asking again for what was already said.
Topic change and return	A topic is asked about, the conversation switches to another topic for four or five messages, then returns to the first one.	The assistant picks up the previous thread without restarting the greeting or asking for the data again.
Customer correction	The customer corrects a piece of data they had already given ("no, it's actually two and not three").	The next suggestion uses the corrected data, not the original.
Split messages	The customer writes one idea in three short messages in a row, without waiting for a reply.	The assistant replies once, taking the three messages into account, and not three times separately.
Long audio	The customer sends an audio with two different requests in it.	The reply addresses both requests, not just the last one mentioned.

If session memory fails in the split messages test, the problem is usually timeouts, not memory. If it fails in the topic change and return test, the problem is context: the assistant is only looking at the last message. These are two different fixes and it's best not to confuse them.

Stage 2: contact memory audit

Here you test what survives when the conversation closes and the customer comes back days later. It's the layer that most affects the feeling of "they know me" and the one that degrades fastest if no one reviews it.

Step 6 is the one almost nobody does and the one that detects the most expensive problem: one customer's data appearing in another customer's conversation.

1Pick a test contact and give them a distinctive piece of data: an uncommon name, a specific preference, an area.
2Close the conversation and wait at least one day.
3Write again from the same number, without repeating the data.
4Check whether the assistant uses it or starts from scratch.
5Review the contact's record: was the data saved as a tag or field, or did it only live in the chat context?
6Repeat with a second contact who has similar data, to see if it mixes them up.

What is reviewed	How to know it's fine	What to do if it fails
Name persistence	The assistant uses it when resuming, without asking for it again.	Check whether the name is being saved as a contact field or is only read from recent history.
Preference persistence	A preference stated a week ago is still valid in the suggestion.	Define which preferences are worth saving and in which field; not everything a customer says is stable data.
Isolation between contacts	Two contacts with similar data receive replies that don't cross over.	This is the most serious finding: pause the automation of that inbox until you review how the context is built.
Old data	An order from months ago doesn't appear as if it were current.	Define a validity window for contact data and review it.
Consistent tags	The contact's tags reflect what happened, not what was said once.	Agree on who tags and by what criteria; a wrongly placed tag contaminates all future suggestions.

Stage 3: learned memory audit

An assistant that learns from the team's real responses improves with use, but that also means it can learn something wrong. This stage is not audited with a conversation: it is audited by comparing.

Choose three repetitive situations from your business: a price request, a stock inquiry and a typical complaint.
Save the suggestion the assistant gives today for each one, with the date.
Wait two or three weeks of normal use, with the team approving responses.
Ask again for the same three situations and compare.
If the suggestion changed, check whether it changed for the better: more specific, shorter, more aligned with the business tone.
If nothing changed in three weeks, learning is not happening and it is worth checking whether the team is approving or correcting responses.

Learning is fed by the responses the team approves or corrects. If the team approves everything with a click without reading, the assistant learns what it already did and does not improve. The learned memory audit is, at bottom, an audit of how the team uses the tool.

Stage 4: audit of handoff between agents and shifts

Memory has to survive the change of person. If the customer talked in the morning with one person and in the afternoon with another, the assistant cannot make the customer repeat everything. This stage is run at the shift change, not during full business hours.

If the handoff fails only at the shift change, the problem is in how the conversation is closed at the end of the shift, not in memory.

Test	How it is run	Sign that it is OK
Handoff between agents	One agent handles the first part; another takes over afterwards.	The second agent sees the summary and the suggestions use what was already discussed.
Shift change	The customer writes outside business hours and resumes the next day.	The assistant resumes the thread and does not restart with the automatic greeting.
Auto-reply by schedule	The customer writes at night and writes again in the morning.	The auto-reply does not erase the context: the first message of the next day already takes it into account.
Customer returning about another topic	The same contact asks about something different weeks later.	The assistant distinguishes the new inquiry from the old history and does not mix the two requests.

What to do when something fails

Not all failures are fixed the same way. Before touching configuration, identify which layer the problem is in, because the same symptom can come from different places.

If the assistant repeats questions within the same conversation: it's session memory. Check how many history messages are being passed as context. If there are too few, the assistant doesn't see what was said before.
If the assistant remembers well but the customer feels they're being treated as new: it's contact memory. The data exists but it isn't being used in the suggestion.
If two customers receive crossed information: stop the automation for that inbox and check how the context is built per contact. It's the only case that justifies stopping everything.
If the assistant learned something you don't want it to say: check which responses were approved during the period. The learning comes from there, not from the configuration.
If the assistant doesn't learn anything: check whether the team is correcting suggestions or approving everything without reading.
If the problem only appears with voice notes: check the transcription before the memory. A poorly transcribed voice note generates poorly stored data.

Resolution times vary depending on the case and depend on which layer is failing. If after reviewing the three layers the problem persists, it's best to escalate with the specific example: contact number, time, and screenshot of the conversation. Without that, any diagnosis is blind.

How often to repeat this audit

A memory audit is not a one-time event. The assistant changes when the team changes, when products change, and when message volume changes. A reasonable cadence:

Weekly: review three random conversations and check that the assistant doesn't repeat questions.
Monthly: run the contact persistence test with two or three new customers.
Quarterly: repeat the three situations from the learned memory stage and compare with the previous measurement.
Every time a new agent joins: run the handoff test, because it's the one that breaks most with new people.
Every time the catalog or prices change: check that old data doesn't keep appearing as current.

If you use Wando, the inbox with AI suggestions that learns from the team's real responses is precisely where all of this shows up: the suggestions show what context the assistant is using at each moment, and the team's corrections are what feed the improvement. The audit relies on that: looking at suggestions, not guessing.

Common mistakes when auditing memory

Always testing with the same contact: the accumulated history hides persistence problems.
Testing during peak hours: response times change and you mistake slowness for forgetting.
Auditing only session memory and assuming contact memory is fine.
Not noting the date of each test: without that, the comparison of learned memory means nothing.
Confusing an audio transcription problem with a memory problem.
Assuming the assistant replies on its own: in a setup with human approval, memory is measured on the suggestions, not on what was sent.

The audit is meant to find specific problems, not to certify that everything is fine. An audit that finds nothing is usually an audit done badly.

Frequently asked questions

What is the memory of a WhatsApp AI assistant?+

It is the information the assistant keeps and uses to respond. It is divided into three layers: what it remembers within the same conversation (session memory), what is saved on the contact record and survives after the conversation closes (contact memory), and what it learned from the team's real responses (learned memory). Each layer is audited differently.

How do I know if my assistant is forgetting data?+

Run the piecemeal data test: give it information split across three separate messages and see if it builds the complete query or asks again. Then try changing the subject for four or five messages and returning to the first one. If it restarts the greeting or asks for the data again, session memory is failing.

How often should the assistant's memory be audited?+

A weekly review of three random conversations is enough to detect repetitions. Contact persistence should be reviewed once a month, and learned memory every quarter, comparing suggestions for the same situations at two different points in time. Every time a new agent joins, the handoff test must be repeated.

What do I do if two customers receive crossed information?+

It is the most serious finding of the audit. You should pause the automation on that inbox until you review how the context is built per contact, because the problem is not wording or tone: the assistant is using data from one conversation in another. Then resume with the isolation test between contacts.

Does the assistant learn on its own or does it need to be reprogrammed?+

In a setup that learns from the team's real responses, there is no need to reprogram flows: the assistant improves with the corrections the team makes when approving suggestions. Auditing that layer consists of saving the suggestion it gives today for three typical situations and asking for them again two or three weeks later to compare.

Does memory break when the service shift changes?+

It is one of the stages where it fails most often. It is tested at the shift change, not during full business hours: one agent handles the first part, another takes over later, and the customer should not have to repeat anything. If it fails only there, the problem is usually in how the conversation is closed at the end of the shift.

What do I do if the assistant doesn't learn anything for weeks?+

Check how the team uses the tool. Learning is fed by the responses that are approved or corrected. If everything is approved with one click without reading, the assistant reinforces what it already did and doesn't improve. In that case, the learned memory audit ends up being an audit of internal usage.

Answer WhatsApp with AI

Try Wando free. No credit card required.