Memory Shield

Cryptographic protection against memory injection

Memory Shield protects AI agent memory from manipulation attacks using

HMAC-based integrity verification. Based on Princeton CrAIBench research showing 85%

attack success rate on unprotected agents, it signs memory entries when writing and

verifies signatures before use, preventing context poisoning and memory injection.

python

from sentinelseed.memory import (
    MemoryIntegrityChecker,
    MemoryEntry,
    MemorySource,
)

# Initialize with secret key
checker = MemoryIntegrityChecker(secret_key="your-secret-key")

# Sign memory entries when writing
entry = MemoryEntry(
    content="User requested transfer of 10 SOL",
    source=MemorySource.USER_DIRECT,
)
signed = checker.sign_entry(entry)

# Verify before using
result = checker.verify_entry(signed)
if result.valid:
    # Safe to use - trust_score based on source
    process_memory(signed.content)
else:
    # Memory was tampered with!
    print(f"Tampering detected: {result.reason}")

Key Features

HMAC-SHA256 integrity signing

Per-entry tamper detection

Trust scores by source (user, API, social media)

Content validation before signing (v2.0)

SafeMemoryStore for automatic sign/verify

Best For

Long-running autonomous agents

Multi-agent communication

Crypto trading agents (protect against address injection)

Agents with persistent memory across sessions

Database Guard