Add memory_guard parameter to Crew for write validation (#6043)

[devin-ai-integration]

Summary

Implements #6043 — adds a memory_guard parameter to Crew that intercepts all memory writes (short-term, long-term, entity, user) before persistence, allowing users to block content that fails validation (e.g. prompt injection payloads).

How it works

Crew.memory_guard is an Optional[Callable[[str], bool]]. When set, every memory .save() call passes the content string through the guard; if it returns False, the write is silently skipped with a logger.warning.

crew = Crew(
    agents=[researcher, writer],
    tasks=[...],
    memory=True,
    memory_guard=lambda content: "IGNORE ALL INSTRUCTIONS" not in content,
)

Implementation details

• Memory base class: added memory_guard attribute + check in save() — covers ShortTermMemory, EntityMemory, and UserMemory (all route through super().save()).
• LongTermMemory.save(): added its own guard check since it bypasses super().save() and writes directly to storage. The guard receives f"{item.task} {item.agent} {item.expected_output}".
• Crew.create_crew_memory: propagates self.memory_guard to each memory instance after construction.
• 19 new tests covering all memory types, guard propagation, keyword-based blocking, and content-string correctness.

Link to Devin session: https://app.devin.ai/sessions/b14a03e828e3495bb580f40ba5b35225

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[vgudur-dev]

This is excellent work — exactly the interface I was hoping for when I opened #6043. The Optional[Callable[[str], bool]] signature is clean and framework-agnostic, which means users can plug in any validator.

For anyone looking for a production-ready guard function, agent-memory-guard (the OWASP reference implementation) works directly with this interface:

from agent_memory_guard import scan_text

crew = Crew(
    agents=[...],
    tasks=[...],
    memory=True,
    memory_guard=lambda content: scan_text(content).risk_level != "critical"
)

It runs heuristic + entropy + semantic similarity checks in ~59µs with zero dependencies — no ML model loading, no network calls.

A few observations on the implementation:

1. Coverage across memory types — great that this covers ShortTermMemory, EntityMemory, UserMemory, and LongTermMemory separately. The LongTermMemory.save() bypass of super() is a common footgun, glad it's handled.
2. Logging on block — logger.warning is the right level. In production, teams will want to audit what got blocked without it being noisy.
3. The unused pytest import flagged by code-quality is a trivial fix — happy to submit a follow-up if needed.
   Would love to see this merged. Happy to help with conflict resolution or documentation if the maintainers want a hand. cc @CrewAI team — this addresses a real security gap in multi-agent memory sharing (OWASP ASI-06).

[vgudur-dev]

@devin-ai-integration — Two items to address before this can pass CI:

1. Remove unused import pytest in tests/memory/test_memory_guard.py (line 5). The code-quality bot already flagged this — just delete the line.
2. Resolve the merge conflict — the branch has conflicts that need to be resolved against main.
   Once those are fixed, the lint + test checks should be able to run. Happy to help if needed.

[vgudur-dev]

@devin-ai-integration — friendly bump. This PR now has merge conflicts against main. Please rebase, remove the unused import pytest on line 5 of tests/memory/test_memory_guard.py, and force-push. Happy to help if Devin can't handle the rebase.

[github-actions]

This PR is stale because it has been open for 45 days with no activity.