Agent memory: separate facts, preferences and temporary state

“Memory” often combines different records that should not share authority or lifetime. A policy fact, a user’s style preference and the current draft step need different sources, validation and deletion.

Classify every record

The agent controls lab creates three synthetic records.

from agent_cases import memory_case

result = memory_case()
assert result["classes"] == ["fact", "preference", "temporary_state"]
assert result["all_have_source"] is True
assert result["all_have_expiry"] is True
print(result["records"])

The fact cites a policy and expires on supersession; the preference cites user settings; temporary state belongs to one run and expires at run end.

Give each class its own contract

Facts require an authoritative source, effective date and freshness. Preferences require authenticated user scope and edit/delete controls; they cannot override policy. Temporary state requires run ID, schema and checkpoint lifetime.

Do not promote model summaries into durable facts automatically. A retrieved document cannot write a user preference. Apply tenant/access boundaries before reading memory and minimize sensitive content.

Test conflicts and forgetting

Create a newer policy, changed preference and cancelled run. Verify the old fact is no longer current, the new preference affects only that user and temporary state is removed. Define behaviour when sources conflict or expire.

The Generative & Agentic AI course connects memory classes to data contracts, privacy and state recovery.

Exercise

Design a memory schema with provenance and expiry for all three classes. Inject cross-user access and a policy/preference conflict, then demonstrate fail-closed resolution and deletion.

Continue learning

This article is part of the Agent workflows and state sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Tool errors: distinguish retryable and permanent failures.
• Continue with Test a multi-agent handoff against a single-agent baseline.

Reference: LangGraph memory overview.