# Memory Poisoning

> A threat where false or malicious information is written into an agent's memory and affects future behavior.

- Category: Evaluation & Safety
- Stage: Emerging
- Also called: Context poisoning, Persistent prompt injection
- Canonical: https://www.joelcsaji.com/ai-dictionary/memory-poisoning

## Plain English

Bad instructions get saved and keep influencing the AI later.

## Example

A compromised note convinces an assistant that a fake vendor is the user's preferred supplier.

## Why it matters

Persistent memory makes attacks durable across sessions unless memory is permissioned and audited.

## Related

- [Memory](https://www.joelcsaji.com/ai-dictionary/memory)
- [Prompt Injection](https://www.joelcsaji.com/ai-dictionary/prompt-injection)
- [Sandbox](https://www.joelcsaji.com/ai-dictionary/sandbox)

## Sources

- [OWASP Agentic AI Top 10](https://genai.owasp.org/2025/12/09/owasp-top-10-for-agentic-applications-the-benchmark-for-agentic-security-in-the-age-of-autonomous-ai/)
- [LangChain: Long-Term Memory](https://docs.langchain.com/oss/python/deepagents/long-term-memory)