memU

Least-privilege permissions

• Scope memory API credentials to specific agents and users; never share a single token across multiple agents or tenants.
• Enforce role-based access control (RBAC) with separate read, write, and delete permissions; restrict agents to only the memory operations they require.
• Rotate API credentials for memory access every 30–90 days and immediately revoke tokens when agents are decommissioned.
• Require explicit consent and audit logs before any agent can access another user's memory store.

Network isolation

• Confirm memory storage is encrypted in transit (TLS 1.2+) and at rest (AES-256 or equivalent); request security documentation from the provider.
• Verify that memory storage is geographically isolated per user or organization; confirm data residency compliance before deployment.
• Implement network segmentation so memory services are only accessible from authorized agent infrastructure; disable public internet access.
• Use VPNs or private endpoints (e.g., AWS PrivateLink) for all memory API calls from agents to backend storage.

Monitoring

• Enable full audit logging for all memory operations (create, read, update, delete) with timestamps, agent IDs, and user identifiers; retain logs for ≥90 days.
• Set up real-time alerts for suspicious patterns: bulk memory access, cross-user queries, credential reuse, or modifications from unexpected agents.
• Regularly review memory access logs and cross-reference with agent behavior; flag anomalies such as agents accessing memories outside their assigned scope.
• Implement integrity checks (e.g., checksums or digital signatures) to detect unauthorized memory modifications.

Red flags to watch for

• Memory storage location is undocumented, vague, or changes without notice; demand explicit confirmation of where data is persisted.
• A single credential or token grants access to multiple users' or agents' memories; this indicates weak isolation.
• No audit trail exists for memory creation, modification, or deletion; this prevents forensic investigation and compliance verification.
• An agent successfully injects false or malicious data into memory and it influences downstream agent decisions; indicates insufficient input validation or memory integrity checks.