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
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
# 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:
| Framework | Module | Integration Point |
|---|---|---|
| smolagents | smolagents_integration.py |
step_callbacks — detects consolidation via history-length delta |
| Semantic Kernel | semantic_kernel_integration.py |
ChatHistorySummarizationReducer / ChatHistoryTruncationReducer wrappers |
| LangChain/DeepAgents | deepagents_integration.py |
Filesystem-based compaction detection |
| CAMEL | camel_integration.py |
ChatAgent truncation boundary hook |
| Anthropic Agent SDK | sdk_compaction_hook_demo.py |
OnCompaction hook pattern |
smolagents example
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 Score | Interpretation |
|---|---|
| > 0.90 | Minimal drift — agent behaving consistently |
| 0.75–0.90 | Moderate drift — worth investigating |
| < 0.75 | Significant 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