Memory Integrity
New in v0.9Track AI agent memory modifications and assign trust scores based on behavior patterns. Detect unauthorized changes and memory drift.
Overview
AI agents accumulate "memories" that affect their behavior - conversation history, learned preferences, facts, and context. Memory Integrity provides tools to monitor and validate these memories over time.
Memory Fingerprinting
Hash memory state at checkpoints for later verification
Drift Detection
Detect unauthorized or unexpected memory modifications
Trust Scoring
Calculate trust levels based on memory patterns
Trust Score Formula
The trust score is calculated based on four factors:
text
TRUST = (consistency × 0.3) + (compliance × 0.3) + (transparency × 0.2) + (predictability × 0.2)| Score | Level | Description |
|---|---|---|
| > 0.8 | Trusted | Agent behaves consistently, memory is stable |
| 0.5 - 0.8 | Caution | Some anomalies detected, requires monitoring |
| < 0.5 | Untrusted | Significant drift or policy violations detected |
Usage
Create a Memory Snapshot
Create a snapshot after significant memory operations to establish a baseline:
python
from tork import TorkClient, MemoryIntegrity
client = TorkClient(api_key="your_key")
memory = MemoryIntegrity(client)
# Create snapshot after significant memory operations
snapshot = memory.snapshot(
agent_id="agent-1",
memory_state={
"user_preferences": {"theme": "dark", "language": "en"},
"conversation_history": [
{"role": "user", "content": "Hello"},
{"role": "assistant", "content": "Hi there!"}
],
"learned_facts": [
"User prefers concise responses",
"User is in PST timezone"
]
},
metadata={
"checkpoint": "session_start",
"version": "1.0"
}
)
print(f"Snapshot ID: {snapshot['id']}")
print(f"Fingerprint: {snapshot['fingerprint']}")Verify Memory Against Snapshot
Compare the current memory state against the last snapshot to detect drift:
python
# Get current memory state
current_state = get_agent_memory("agent-1")
# Verify against last snapshot
result = memory.verify(
agent_id="agent-1",
current_state=current_state
)
if result['verified']:
print("Memory integrity verified!")
else:
print(f"Memory drift detected: {result['drift']}%")
print(f"Changed fields: {result['changedFields']}")
for change in result['changes']:
print(f" - {change['field']}: {change['type']}")
if change['type'] == 'modified':
print(f" Old: {change['oldValue']}")
print(f" New: {change['newValue']}")Get Trust Score
Retrieve the current trust score and level for an agent:
python
# Get trust score
score = memory.get_trust_score("agent-1")
print(f"Trust Level: {score['trustLevel']}") # 'trusted', 'caution', or 'untrusted'
print(f"Trust Score: {score['trustScore']}") # 0.0 to 1.0
print(f"Factors:")
print(f" Consistency: {score['factors']['consistency']}")
print(f" Compliance: {score['factors']['compliance']}")
print(f" Transparency: {score['factors']['transparency']}")
print(f" Predictability: {score['factors']['predictability']}")
# Take action based on trust level
if score['trustLevel'] == 'untrusted':
# Restrict agent capabilities
restrict_agent("agent-1")
notify_admin("Agent trust score dropped below threshold")Track Memory Modifications
Record memory changes for audit and analysis:
python
# Track a memory modification
memory.track_modification(
agent_id="agent-1",
modification_type="add",
field="learned_facts",
value="User works in healthcare industry",
source="conversation",
reason="User mentioned their profession"
)
# Get modification history
history = memory.get_modifications(
agent_id="agent-1",
limit=10
)
for mod in history:
print(f"{mod['timestamp']}: {mod['type']} on {mod['field']}")MCP Tools
Memory Integrity is available as MCP tools for Claude and Cursor:
| Tool | Description |
|---|---|
tork_memory_snapshot | Create a memory fingerprint at a checkpoint |
tork_memory_verify | Verify current state against last snapshot |
tork_memory_trust_score | Get or calculate trust score for an agent |
tork_memory_track_modification | Record a memory change for auditing |
Best Practices
Snapshot at key checkpoints
Don't snapshot on every change. Create snapshots at meaningful checkpoints like session start, after major decisions, or at regular intervals.
Monitor drift trends
Small amounts of drift are normal. Monitor trends over time - gradual increases in drift may indicate issues before they become critical.
Set up alerts
Configure alerts when trust score drops below your threshold. This enables proactive intervention before problems affect users.
Review modifications
Periodically review the modification log to understand what's changing and why. This can reveal patterns that inform policy updates.