Q&A: Memory, State & Context — How CMMC Agents Remember & Persist Between Sessions

[AGI-Corporation]

Memory, State & Context Deep-Dive Q&A

This thread is for deep technical questions about how the CMMC AI agents manage memory, persist state between runs, and maintain context across long compliance workflows.

---

Common Questions

Session memory & context windows

• How much context does each agent maintain per session (token limits, sliding windows)?
• How does the assessment agent remember which controls it already evaluated in a long session?
• What happens to agent memory when a session times out or crashes?
• Can I configure context retention per agent type?

Persistent state & database

• What data is persisted to the database between agent invocations?
• How is assessment state stored — per-control, per-domain, or per-session?
• How do I inspect or export the current state of an ongoing assessment?
• Can state be rolled back to a previous checkpoint?

Long-running assessments

• How does the platform handle a multi-day assessment that spans multiple sessions?
• Is there a resume/checkpoint mechanism for interrupted assessments?
• How are partial results saved so work isn't lost on failure?

Multi-agent state sharing

• How do the assessment, remediation, and report agents share state?
• Is there a shared memory bus or message queue between agents?
• How are race conditions handled when multiple agents update the same control record?

Vector memory & semantic search

• Does any agent use a vector store for semantic retrieval (e.g., finding similar controls or past evidence)?
• Which embedding model is used for semantic memory, and how is it configured?
• How do I add custom knowledge (policy docs, org-specific guides) to the agent's retrieval context?

---

Share your memory architecture questions, state management challenges, or ideas for improving agent persistence below.