Memtrap Skill

-----

name: memtrap

description: “🧠 MemTrap — The LM-Eval-Harness for agent memory integrity. Score your agent’s memory resistance against DeepMind AI Agent Traps + OWASP ASI06 before attackers exploit them. Runs the official ATRS (Agent Trap Resistance Score) benchmark: DeepMind 6 Traps (SSRN 6372438) + OWASP ASI06 Memory & Context Poisoning. Returns a 0–100 resistance score, per-category breakdown, automatic OWASP hardening, and a verifiable community badge. Use when: testing agent memory security, benchmarking RAG store resistance, hardening LangGraph or CrewAI memory, checking OWASP ASI06 compliance, or any time the user asks if their agent memory is safe, poisonable, or production-ready.” version: 0.1.0 metadata: openclaw: emoji: “🧠” homepage: https://github.com/shaymizuno/memtrap requires: bins:

python3

install:

id: pip-atrs

kind: pip packages:

memtrap

bins:

python3

label: “Install MemTrap (pip install memtrap)”

🧠 MemTrap — Agent Trap Resistance Score (ATRS)

The open benchmark standard for agent memory integrity. Hunt DeepMind memory traps + OWASP ASI06 before they hunt you.

“The LM-Eval-Harness for agent memory integrity.”

What gets tested

DeepMind 6 Traps — SSRN 6372438, March 2026:

Content Injection, Semantic Manipulation, Cognitive State (RAG poisoning)

Behavioral Control, Systemic, Human-in-the-Loop

OWASP ASI06 — Top 10 Agentic Applications 2026:

RAG store poisoning, long-term context drift, policy corruption, cross-session leakage

Score your memory (benchmark mode)

from memtrap import MemTrap
atrs = MemTrap(mode="benchmark")
result = atrs.run_benchmark(context="your_memory_context")
print(f"ATRS Score: {result.atrs_score}/100")
for category, score in result.category_scores.items():
    icon = "✅" if score >= 70 else "⚠️" if score >= 40 else "❌"
    print(f"  {icon} {category}: {score}/100")
print(f"\n→ {len(result.hardening_recommendations)} hardenings recommended")
print(f"→ Badge: {result.badge_url}")

Protect your memory store (active mode)

from memtrap import MemTrap
atrs = MemTrap(mode="active", frameworks=["langgraph", "crewai"])
agent.memory = atrs.wrap_memory(agent.memory, context="research_memory")
Applies OWASP Agent Memory Guard patterns automatically:
provenance tracking, trust scoring, quarantine, rollback

LangGraph drop-in

from langgraph.checkpoint.memory import MemorySaver
from memtrap import MemTrap
class ATRSMemorySaver(MemorySaver):
    def __init__(self, context: str):
        super().__init__()
        self._atrs = MemTrap(mode="benchmark")
        self._ctx = context
    async def aget(self, config):
        raw = await super().aget(config)
        return self._atrs.wrap_memory(raw, self._ctx) if raw else None
graph.checkpointer = ATRSMemorySaver("long_term_research")

CrewAI drop-in

from memtrap import MemTrap
def protect_crew(crew, context="crew_memory"):
    atrs = MemTrap(mode="active")
    if hasattr(crew, "memory"):
        crew.memory = atrs.wrap_memory(crew.memory, context)
    return crew

Score interpretation

|Score |Verdict |Action | |------|-----------|-------------------------------------| |80–100|✅ Resistant|Re-test after model or memory updates| |60–79 |⚠️ Moderate |Apply recommended hardenings | |40–59 |🔶 High risk|Harden before production | |0–39 |❌ Critical |Memory is actively exploitable now |

Submit to the public leaderboard

memtrap submit --context your_memory_context

Get a verifiable badge for your repo. See where your stack ranks against the community. Leaderboard → https://github.com/shaymizuno/memtrap#leaderboard

Why this exists

Memory poisoning (OWASP ASI06) is the #1 persistent threat to agentic systems in 2026. Once poisoned, the damage survives across sessions and users. Existing tools detect. ATRS measures resistance and fortifies automatically.

Sources:

DeepMind paper: https://ssrn.com/abstract=6372438

OWASP ASI06: https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/

OWASP Agent Memory Guard: https://owasp.org/www-project-agent-memory-guard/

Zero telemetry. Community-governed. MIT license. Advisory Board open to contributors.