[FEATURE] Memory write guards for multi-agent crews — prevent cross-agent memory poisoning

Feature Area

Core functionality

Is your feature request related to a an existing bug? Please link it here.

NA - not related to an existing bug

Describe the solution you'd like

In multi-agent CrewAI crews, agents share memory (short-term, long-term, entity memory). A compromised or manipulated agent can poison shared memory with prompt injection payloads that affect all other agents in the crew.

Proposed solution: Add a memory_guard parameter to Crew that validates all memory writes before persistence:

from crewai import Crew, Agent, Task
from agent_memory_guard import scan_text

def memory_validator(content: str) -> bool:
    result = scan_text(content)
    return result.risk_level != "critical"

crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[...],
    memory=True,
    memory_guard=memory_validator  # New parameter
)

This would intercept writes to all memory types (short-term, long-term, entity) and block content that matches injection patterns. The agent-memory-guard package (OWASP project, 59µs latency, zero deps) provides a production-ready validator.

Why this matters for CrewAI specifically: Multi-agent crews amplify the risk because one poisoned memory entry propagates to ALL agents. A single injection in a researcher agent's output can compromise the writer and reviewer agents permanently.

Describe alternatives you've considered

1. Per-agent input validation — doesn't catch tool outputs or inter-agent messages that flow into memory
2. Post-hoc memory cleanup — reactive rather than preventive, damage already done by the time you detect it
3. Disabling shared memory — defeats the purpose of multi-agent collaboration

Additional context

• OWASP ASI-06 (Memory Poisoning): https://owasp.org/www-project-agentic-security-initiative/
• Reference implementation: https://pypi.org/project/agent-memory-guard/
• GitHub: https://github.com/OWASP/www-project-agent-memory-guard
• I have a working prototype of this integration and would be happy to submit a PR

Willingness to Contribute

Yes, I'd be happy to submit a pull request

Start at the Crew entry point and trace how short-term, long-term, and entity memory writes are persisted. Review the proposed memory_guard callback and the prototype or agent-memory-guard reference, then define coverage for every memory type. Done means critical-risk content is rejected before persistence without disabling shared memory.