External memory is information stored outside an AI model that the system can read again later, such as a database, file, wiki, cache, or task log. Unlike the model’s built-in parameters, this content can be updated, shared, or reused across sessions and runs.
In cybersecurity, external memory matters because it extends the attack surface. If the storage is writable, an attacker may poison it with misleading instructions, malicious links, or corrupted context, causing later runs to behave incorrectly. If it is readable without proper controls, sensitive data can leak into future prompts or agents. Defenders treat external memory as untrusted by default, separate data from instructions, verify provenance, restrict write access, and log reads and updates. In agentic systems, external memory can improve continuity, but it can also turn persistence into a security risk.