LangMem — agentic threat model

Not certain from the listing — LangMem is model-agnostic and does not specify foundation model requirements, though its prompt optimization features directly influence how downstream LLMs interpret instructions and handle adversarial inputs.

LangMem directly manages long-term memory and conversation history. This introduces significant risks of memory poisoning, where malicious conversation inputs are permanently extracted into the knowledge base, and data exfiltration if sensitive user data is stored without proper encryption or access controls.

As an agent memory framework, LangMem's background memory manager and prompt optimization primitives are susceptible to indirect prompt injection. Attackers can manipulate conversation history to alter the agent's optimized prompts, leading to unauthorized tool execution or persistent behavioral manipulation.

Not certain from the listing — LangMem is an open-source toolkit, meaning deployment security, sandboxing, and database access controls are entirely dependent on the user's infrastructure implementation.

Not certain from the listing — The toolkit does not detail built-in guardrails, anomaly detection, or monitoring tools to detect when memory extraction has been compromised or poisoned.

Not certain from the listing — There is no mention of built-in authentication, authorization, or compliance frameworks (such as GDPR/CCPA data deletion compliance for long-term stored memories).

Not certain from the listing — While it integrates with LangGraph (which supports multi-agent architectures), the listing does not specify how trust boundaries or memory access permissions are managed between different agents sharing the same storage layer.