Separate assistant memory from customer truth
Assistant memory should distinguish temporary conversation context, user-confirmed preferences, and authoritative business facts. Store provenance and scope with remembered items, refresh customer facts from their owning systems, and make corrections invalidate derived summaries rather than allowing old statements to become permanent truth.
Inside this article 5 sections +
An assistant remembers that a customer is “on the annual plan” because the customer said so last month. The CRM now shows a monthly plan, but a conversation summary still contains the old statement. Without an authority model, the assistant may treat the remembered sentence as more convenient evidence than the current account record.
Memory architecture should answer three questions for every stored item: who established it, where it applies, and what can make it invalid. A longer context window does not resolve those questions; it merely keeps more potentially conflicting statements available.
Use different stores for different kinds of memory
Conversation context helps interpret references such as “that report” within an ongoing task. User preferences describe choices such as preferred language or response length. Business facts describe account state, entitlements, contacts, and transactions owned by application systems. These categories should not share one unrestricted fact table.
Temporary task references and working assumptions, tied to the conversation.
User-confirmed choices with scope, correction, and optional expiry.
Authoritative account facts fetched from the owning service.
Keep the category explicit in storage and in context assembly. A model-generated summary should not promote an uncertain user statement into a verified entitlement. If the task needs the current plan, fetch it from the account service and label its source and observation time.
Use a conflict rule: authoritative current facts govern business decisions, while a conflicting conversation statement becomes something to clarify. The assistant can say that the account record differs from the earlier discussion without silently rewriting the customer record to match the conversation.
Make remembered preferences inspectable
Store a preference's subject, scope, value, origin, confirmation status, and update time. “Use concise answers” may apply to one user; “send account notices in French” may require an account-level setting and appropriate authority. A shared company account should not inherit one employee's personal preference by accident.
{
"subject_id": "user-demo-8",
"scope": "assistant_response",
"key": "preferred_language",
"value": "French",
"origin": "explicit_user_request",
"confirmed": true,
"revision": 3,
"supersedes_revision": 2
}
This example is an application memory record. It does not authorize changing the CRM's communication settings or sending messages. Those are separate business actions with their own contracts and permissions.
Give users a way to inspect, correct, and remove remembered preferences. Avoid an interface that only offers “forget everything” when one item is wrong. Store corrections as new revisions or otherwise preserve enough lineage to invalidate dependent summaries and caches.
Choose expiry by meaning, not convenience
A temporary project reference may expire when the task closes. A language preference may remain until changed. A cached account balance may require a fresh read for every consequential decision. One global memory time-to-live cannot express all three policies.
Redis expiry or another cache mechanism can implement storage lifetime, but it does not establish whether a remembered fact is still valid. Keep business freshness requirements separate from physical key expiration. A cache entry can exist and still be ineligible for the current task.
For facts fetched from an owning system, retain the source identifier, revision, and observed time. Decide which tasks may use a cached value and which require revalidation. Reading an account label for conversational context may tolerate different staleness from deciding whether a customer may access a paid feature.
Handle unknown values explicitly. If the account service is unavailable, a remembered statement should not silently become the fallback truth. Return an unavailable or needs-verification state, or use a documented stale-data policy whose limitations the product can explain.
Follow a correction through derived memory
Suppose a user changes their preferred language from English to French. Updating the preference row is insufficient if a conversation summary says “always respond in English” and a cached context plan keeps reintroducing it. Track which summaries and plans depend on the old revision.
Invalidate or regenerate those derivatives when the preference changes. During context assembly, prefer the latest confirmed revision and exclude superseded preference statements from model instructions. Historical messages can remain as records without continuing to control present behavior.
For a business-fact correction, update the owning system through its normal workflow and refresh the assistant's derived state. Do not let a conversational correction write arbitrary profile fields. “Actually, I am the account owner” is a claim to verify, not a permission grant.
Generated summaries should preserve uncertainty. If the user says they may move offices next quarter, memory should not store the proposed address as the current delivery address. A useful summary distinguishes plans, claims, preferences, and verified facts.
Test memory over multiple sessions
Create a synthetic sequence with a stated preference, a later correction, a changed account plan, a conversation summary, and a new session. Verify that the assistant preserves useful continuity while using the current preference and authoritative plan. Inspect the assembled context to find stale derivatives even if the final answer looks correct.
Repeat with a shared account, a user switching tenants, and a deleted source. Memory scope and authorization must remain aligned. A personalization feature should not become a cross-account shortcut around the retrieval boundary.
Measure correction success, stale-memory reuse, unnecessary refetches, and user effort to fix a remembered item. More stored facts are not automatically better memory. The useful measure is whether continuity reduces effort without creating unsupported assumptions.
Read chat-history access control for revocation and knowledge freshness for source age. Our AI integration work connects those lifecycle rules to customer-facing assistants.