memU — agentic threat model

Not certain from the listing — memU is a memory infrastructure layer for LLM applications rather than a foundation model itself, though its context injection directly influences model outputs and could be abused to trigger model alignment failures.

Highly critical layer for memU. It manages RAG retrieval, memory graphs, and file-system-like memory storage. Key threats include memory/knowledge-base poisoning, where malicious inputs are permanently stored as 'files' or 'symlinks' and subsequently injected into the agent's context.

Directly relevant as memU provides the memory and context-injection orchestration. Vulnerabilities in the memory graph structure or proactive intent-prediction algorithms could allow attackers to manipulate the agent's planning and decision-making flow.

Involves the deployment of memU-server, memU-ui, and hosted APIs. Threats include unauthorized API access to the Memory/Response endpoints, lack of transport security, and potential host compromise of the self-hosted server instances.

Not certain from the listing — there is no explicit mention of built-in evaluation frameworks, guardrails, or anomaly detection to identify when poisoned or malicious memories are being injected into the context.

Not certain from the listing — the description does not outline specific authentication, authorization (RBAC) for memory access, or compliance standards (like SOC2 or GDPR) for the hosted API service.

Highly relevant for multi-agent systems. If multiple agents share the same memU infrastructure, a compromise or malicious action by one agent could poison the shared memory graph, leading to cascading failures and trust abuse across the entire ecosystem.