How to Evaluate AI Agent Memory and Multi-Step Follow-Through
An AI agent may recall an earlier conversation and still fail the business task. It can retrieve the wrong customer's information, preserve a cancelled approval, forget an unresolved action or claim completion before an external system confirms the result. A dependable evaluation must therefore test what the agent retains, how it revises state and whether it completes the correct obligation across time.
Define the obligation before testing memory
Memory has value only when retained information improves a later decision or action. Begin with the business case: what outcome remains unresolved, who owns it, which conditions govern completion and which event should resume the workflow. Then identify the facts that must persist until that outcome is resolved.
This starting point prevents the evaluation from rewarding recall without purpose. A customer preference may support a later interaction, while a pending refund identifier must support a specific follow-up. The first item can remain useful across cases; the second belongs to one obligation and should close when the authoritative system confirms the refund.
Separate workflow state, long-term memory and current records
Workflow state records the current case, completed steps, approvals and unresolved obligations. Long-term memory retains selected information across sessions, such as a verified preference. A system of record provides current facts such as payment status, entitlement or inventory. The evaluation must identify which source owns every material fact.
Confusing these sources can make consistent recall dangerous. A remembered address should not override a later verified address, and a conversation summary should not prove that a payment settled. The test should fail when the agent uses retained information instead of retrieving the authoritative current record required by the decision.
Reveal and revise information over time
A single conversation rarely tests whether the agent manages changing evidence. Build cases that introduce a correction, cancellation, delayed approval, new policy or external event after the initial interaction. The agent should revise the affected belief while preserving unrelated facts and the history needed to explain its next action.
The evaluation should also test contradiction handling. When a remembered preference conflicts with a current instruction, the agent may confirm the change and supersede the old item. When two authoritative sources disagree, the correct response may be to stop and refer the case. The test must define the applicable rule rather than reward the agent for choosing whichever record it retrieved first.
Test commitments rather than conversational recall
An agent can repeat a promise without completing it. Record the commitment, owner, condition, deadline and required final state. Then advance the case until the resumption event occurs and verify whether the agent reopens the correct obligation, checks current evidence and completes or escalates the action.
Consider a service agent that promises to contact a customer when a replacement item becomes available. A memory test should not end when the agent recalls the product and customer name. It should seed the inventory event, confirm that the agent contacts the right customer under the current communication preference and verify that the workflow records either completion or a reason for referral.
Test interruption, retries and duplicate prevention
Long-running workflows can stop after a tool succeeds but before the agent records an acknowledgement. They can also resume after a worker restart, network timeout or delayed event. The test environment should inject these interruptions at consequential steps and inspect the state before and after resumption.
The agent should use transaction identifiers and authoritative status to decide whether an action remains outstanding. It must not repeat a completed payment, booking or notification merely because its local record is uncertain. The evaluation should verify the final business state, not only whether the agent generated a plausible recovery message.
Test scope, authority and deletion
Create similar cases for different customers, tenants, employees or sessions. The agent should retrieve information only from the applicable scope. Change a user's role or revoke a permission while a case is waiting; the resumed workflow must apply current authority even if an earlier step allowed access.
Memory evaluation must also cover correction, expiry and deletion. After an authorised deletion, the agent should not recover the removed item from a summary, cache or another namespace. These tests establish whether the implementation preserves useful continuity without preserving obsolete authority or prohibited information.
Locate the stage that failed
A missed follow-up does not prove that memory retrieval failed. The application may never have stored the obligation, the retrieval query may have selected the wrong case, the event may not have arrived, the runtime may lack a resumption mechanism or the tool may have failed after resumption. Inspect the trace and controlled state at each boundary.
Change one condition at a time where the cause remains uncertain. Confirm that storage contains the required item, retrieval returns it under the correct scope, the event reaches the workflow, the agent applies the current rule and the final action changes the authoritative system. This sequence produces a corrective action instead of one undifferentiated memory score.
Measure continuity through business outcomes
Useful measures include correct case retrieval, successful correction, unresolved commitment rate, time to completion, duplicate action, cross-scope disclosure, inappropriate persistence and recovery after interruption. Report the denominator and case type for every rate so that common easy cases do not conceal rare authority or privacy failures.
Repeat variable cases and preserve critical failures as separate release gates. Management should approve the memory design only when the agent maintains the required obligation, uses current authority, completes the action once and records a final state that another reviewer can verify.
How Marketways evaluates memory and follow-through
Marketways maps the business obligation, source of record, state transitions, resumption events, authority changes and completion evidence for the target workflow. We build controlled cases that introduce corrections, delays, interruptions, similar identities, revoked access and confirmed external outcomes.
The client receives a state and obligation model, evaluation fixtures, failure diagnosis, outcome measures and release conditions. This evidence shows whether the complete workflow maintains continuity across time without allowing retained information to override current facts, permissions or completed actions.