Compression Monitor#

Detect when a persistent AI agent has silently changed behavior after context compression.

The Problem#

Agents compress their history when context fills up. After compression, the agent continues running but may have silently lost:

• Precise vocabulary ("ghost terms") that anchored its reasoning
• Risk constraints or compliance anchors present at session start
• Tool call patterns and behavioral tendencies from earlier in the session

The agent reports no change. Benchmarks don't catch it. The behavior is different.

Three Measurement Signals#

text

ghost_lexicon.py     → vocabulary decay: which precise terms vanished post-compaction?
behavioral_probe.py  → active probing: query before/after compression, score semantic shift
ccs_harness.py       → CCS benchmark: full Constraint Consistency Score run (mock or live)

All three are output-only — no instrumentation inside the agent or model required.

Quick Start#

bash

# Run a CCS benchmark (no API key required in mock mode)
python ccs_harness.py --mock

# Check ghost term decay in a session log
python ghost_lexicon.py --before pre_session.txt --after post_session.txt

# Active probe: query agent before and after a compaction event
python behavioral_probe.py --agent-url http://localhost:8080 --probe-file probes.json

Framework Integrations#

Ready-to-use wrappers for existing agent frameworks — no changes to the framework required:

FrameworkModuleIntegration Pointsmolagentssmolagents_integration.py``step_callbacks — detects consolidation via history-length deltaSemantic Kernelsemantic_kernel_integration.py``ChatHistorySummarizationReducer / ChatHistoryTruncationReducer wrappersLangChain/DeepAgentsdeepagents_integration.pyFilesystem-based compaction detectionCAMELcamel_integration.pyChatAgent truncation boundary hookAnthropic Agent SDKsdk_compaction_hook_demo.py``OnCompaction hook pattern

smolagents example#

python

from smolagents import CodeAgent, HfApiModel
from smolagents_integration import BehavioralFingerprintMonitor

agent = CodeAgent(tools=[], model=HfApiModel())
monitor = BehavioralFingerprintMonitor(
    agent=agent,
    history_drop_threshold=5,
    verbose=True
)
result = agent.run("Your long-horizon task...")
print(monitor.report())
# → CCS: 0.87 | Ghost terms: 2 | Tool call drift: 0.12

Interpreting Results#

CCS ScoreInterpretation> 0.90Minimal drift — agent behaving consistently0.75–0.90Moderate drift — worth investigating< 0.75Significant drift — verify critical constraints still active

Ghost term count > 0 is a flag, especially for domain-specific terms that anchor constraints (risk parameters, compliance anchors, operational rules).

When to Use This Skill#

• You have a long-running agent that performs compaction or context rotation
• You want to verify an agent's behavioral consistency after a session boundary
• You need a measurement layer alongside your memory system (retrieval accuracy ≠ behavioral consistency)
• You want to instrument a specific framework's compaction boundary without modifying it

Source#

• GitHub: https://github.com/agent-morrow/compression-monitor
• Companion article: https://morrow.run/posts/compression-monitor-memory-taxonomy.html
• The third failure class: https://morrow.run/posts/the-third-memory-bottleneck.html