What is business memory for an AI agent?
Business memory is a governed, persistent record of the knowledge and decisions an agent needs to work for one company over time. It is different from a chat transcript: it stores source documents, operating policies, dated observations, recommendations, user responses and outcome snapshots as explicit business objects.
Why is conversation history not enough?
Conversation history is unstructured, user-specific and difficult to audit. A product team needs to know which policy is current, who supplied a fact, when it became effective, which recommendation was accepted and whether the result supported it. Those records should remain available across sessions, users and interfaces.
What should be stored?
- Company policies and decision thresholds, with owner and effective date.
- Private documents and extracted facts, preserving source and authority.
- Agent recommendations, evidence and confidence.
- The human decision: accepted, modified, rejected or deferred.
- Dated KPI snapshots and, when available, the observed outcome.
How should memory be governed?
Each tenant’s memory should remain isolated. Users need to inspect, correct and remove memories, distinguish supplied policy from inferred pattern and see why a record exists. Sensitive data needs the same access control and retention discipline as the underlying operational system.
What does Cycle support today?
Cycle provides tenant knowledge ingestion and operational memory for policies, sources, recommendations, responses and periodic snapshots. Cycle Open uses a public knowledge layer; connected businesses receive their own context. Outcome-driven confidence adjustment is direction and should not be confused with a fully autonomous learning loop already deployed everywhere.