tracegc-crewai

tracegc-crewai is an official adapter package providing TraceGC graph-based context compaction for CrewAI agents, crews, and tasks.

Installation

pip install tracegc-crewai

Or install with test dependencies:

pip install tracegc-crewai[test]

Quick Start

1. Compact CrewAI Execution Steps via step_callback

Use TraceGCCrewCallback or create_step_callback() to record and compact step outputs during Crew execution:

from crewai import Agent, Crew, Task
from tracegc_crewai import create_step_callback

# Create a TraceGC step callback
step_cb = create_step_callback(
    prune_semantic=True,
    tracked_decision_keys={"database", "cache_backend"}
)

# Instantiate CrewAI Agent and Crew with step_callback
agent = Agent(
    role="Backend Architect",
    goal="Design scalable infrastructure",
    backstory="Senior system architect"
)

task = Task(
    description="Select storage engine and caching layer",
    expected_output="Final tech stack recommendation",
    agent=agent
)

crew = Crew(
    agents=[agent],
    tasks=[task],
    step_callback=step_cb
)

# Run crew...
# crew.kickoff()

# Compact accumulated step context
result = step_cb.compact()
print("Compacted messages:", len(result.messages))
print("Pruned receipts:", len(result.receipts))

2. Manual Context Compaction via compact_messages

If you manage task histories or step outputs directly:

from tracegc_crewai import compact_messages

step_history = [
    "Configured database=postgres",
    "Switched database=mongodb due to schema flexibility",
]

result = compact_messages(step_history, prune_semantic=True, tracked_decision_keys={"database"})
for msg in result.messages:
    print(msg["content"])
# Output retains only the active 'mongodb' decision and prunes the superseded 'postgres' choice.

Why TraceGC instead of CrewAI's Built-in Memory?

CrewAI provides a built-in memory subsystem (memory=True) featuring ShortTermMemory (using ChromaDB vector RAG), LongTermMemory, and EntityMemory.

While CrewAI's built-in memory excels at semantic retrieval (RAG) across tasks:

• CrewAI Memory retrieves relevant facts by embedding similarity, but retains all past execution steps in raw history. It does not prune obsolete variables, superseded tech choices, or dead search branches.
• TraceGC builds a deterministic dependency DAG over trace events. It mathematically prunes obsolete states, superseded configuration decisions, and abandoned branches while preserving exact provenance and recoverable receipts.

Using tracegc-crewai alongside CrewAI's memory ensures your agents run with minimal token context windows, zero obsolete state confusion, and full auditability.