mcal-ai-autogen 0.5.0

Microsoft AutoGen integration for MCAL - Goal-aware memory for multi-agent systems

mcal-ai-autogen

Microsoft AutoGen integration for MCAL (Memory-Context Alignment Layer), bringing goal-aware memory to AutoGen agents.

Installation

pip install mcal-ai-autogen

# With AutoGen dependencies
pip install mcal-ai-autogen[autogen]

Quick Start

from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient
from mcal import MCAL
from mcal_autogen import MCALMemory

# Initialize MCAL
mcal = MCAL(llm_provider="openai")

# Create MCAL-backed memory
memory = MCALMemory(mcal, user_id="user_123")

# Create an agent with MCAL memory
model_client = OpenAIChatCompletionClient(model="gpt-4")
agent = AssistantAgent(
    name="data_engineer",
    model_client=model_client,
    memory=[memory],
    system_message="You are a helpful data engineering assistant.",
)

# Use the agent — MCAL automatically tracks context and decisions
result = await agent.run(task="How should I set up my ETL pipeline?")

What's New in 0.5.0

• Query-Aware Subgraph Retrieval — New seed-and-expand pipeline replaces 6 query-blind retrieval paths with a single query-aware pass. Reduces context tokens by 53% at 1020 turns while improving DRR by 4.5pp.
• QuerySubgraph dataclass — New public API for structured subgraph results, partitioned by node type (goals, decisions, facts, entities, actions) with structural edge resolution.
• Adjacency index — Lazy-built bidirectional adjacency index on UnifiedGraph enables O(1) neighbor lookups for graph traversal.
• Improved DRR at scale — CTO-1020 DRR improved from 85.3% to 89.9% (+4.6pp); CTO-300 improved from 92.2% to 94.4% (+2.2pp).
• LoCoMo-10 Evaluation — Full 10-conversation, 1,540 QA binary evaluation: 46.1% overall accuracy.

What's New in 0.4.1

• First-Class FACT Nodes — 3 new typed edges (measures, evidence_for, quantifies) improve fact retrieval; quantitative queries automatically boost fact content
• Importance Scoring Boost — FACT nodes with numeric values score higher in retrieval
• search_facts() API — Filter facts by category and value range on UnifiedGraph
• Version Metadata Fix — __version__ now correctly reports 0.4.1 (was stuck at 0.2.9)

What's New in 0.4.0

• Graph Compaction Fixes — Improved retrieval quality with facts-in-context, expanded edge types, chunk boost scoring
• CTO-1020 Benchmark — 85.3% decision retention over 1020 turns, 95.6% cross-era recall, 88% token reduction
• Statistical Rigor — Multi-run validation with Fisher's exact test, Wilson score confidence intervals

What's New in 0.3.0

• Expanded Relationship Edge Types — 10 new edge types (family, friend, colleague, likes, prefers, lives_in, works_at, etc.) for richer relationship graphs
• Key Facts & Entities in Search Context — search() now surfaces extracted facts and background entities directly in result.context
• Improved Chunk Retrieval — More results returned with equal weighting; conversation excerpts prioritized in context

Older releases

What's New in 0.2.9

• Configurable Extraction Profiles — Choose decision, conversational, or comprehensive
• Hybrid Retrieval with ChunkStore — Graph traversal + embedding search for maximum recall
• FACT/PERSON Node Protection — Graph compaction preserves factual and identity nodes

# Pass extraction options to MCAL
mcal = MCAL(
    llm_provider="anthropic",
    extraction_profile="decision",
    enable_chunk_store=True,
)
memory = MCALMemory(mcal, user_id="user_123")

Features

Goal-Aware Memory

MCAL's unique value is understanding your project's goals and maintaining context across conversations:

mcal = MCAL(llm_provider="anthropic")
memory = MCALMemory(mcal)

# Add relevant context
from autogen_core.memory import MemoryContent
await memory.add(MemoryContent(
    content="We decided to use Kafka for streaming",
    mime_type="text/plain",
    metadata={"category": "architecture", "decision": True}
))

# Query returns goal-relevant results
results = await memory.query("What messaging system should I use?")
# Returns Kafka decision with goal-relevance scoring

Decision Tracking

Track architectural and project decisions automatically:

memory = MCALMemory(
    mcal,
    enable_goal_tracking=True,  # Extract goals from content
    include_decisions=True,      # Include decisions in search
)

# Decisions are automatically tracked
await memory.add(MemoryContent(
    content="After evaluating options, we chose PostgreSQL for its JSON support",
    mime_type="text/plain"
))

# Query finds relevant decisions
results = await memory.query("database selection")

User Isolation

Support multi-tenant scenarios with user isolation:

# Create separate memories for different users
user1_memory = MCALMemory(mcal, user_id="alice")
user2_memory = MCALMemory(mcal, user_id="bob")

# Each user has isolated memory
await user1_memory.add(MemoryContent(content="Alice prefers Python"))
await user2_memory.add(MemoryContent(content="Bob prefers Rust"))

# Queries only return user-specific results
results = await user1_memory.query("language preference")
# Only returns Alice's preference

TTL Support

Configure time-to-live for memory entries:

memory = MCALMemory(mcal, default_ttl_minutes=60)  # 1 hour default

# Or per-entry TTL via metadata
await memory.add(MemoryContent(
    content="Temporary context",
    mime_type="text/plain",
    metadata={"ttl_minutes": 15}  # 15 minute TTL
))

Thread Safety

All operations are protected by RLock — safe for concurrent access from multiple agents.

Integration with AutoGen Features

With AssistantAgent

from autogen_agentchat.agents import AssistantAgent

agent = AssistantAgent(
    name="assistant",
    model_client=model_client,
    memory=[memory],  # MCAL memory integrates seamlessly
)

With Teams

from autogen_agentchat.teams import RoundRobinGroupChat

# Share MCAL memory across team members
shared_memory = MCALMemory(mcal, user_id="team_alpha")

coder = AssistantAgent("coder", model_client=model_client, memory=[shared_memory])
reviewer = AssistantAgent("reviewer", model_client=model_client, memory=[shared_memory])

team = RoundRobinGroupChat([coder, reviewer])

Context Window Management

MCAL automatically manages context relevance:

memory = MCALMemory(
    mcal,
    max_results=10,           # Limit results per query
    score_threshold=0.5,      # Minimum relevance score
)

# update_context adds relevant memories to the agent's context
result = await memory.update_context(model_context)

API Reference

MCALMemory

class MCALMemory(Memory):
    def __init__(
        self,
        mcal: MCAL,
        user_id: str = "default",
        name: str = "mcal_memory",
        max_results: int = 10,
        score_threshold: float = 0.0,
        default_ttl_minutes: Optional[float] = None,
        enable_goal_tracking: bool = True,
        include_decisions: bool = True,
    ): ...

Key Methods

Method	Async	Description
add(content)	✓	Add MemoryContent to memory
query(query)	✓	Search for relevant memories, returns MemoryQueryResult
update_context(model_context)	✓	Update agent context with relevant memories
clear()	✓	Clear all memory entries
close()	✓	Cleanup resources

Helper Methods

Method	Description
add_text(text, metadata=None)	Convenience wrapper for adding plain text
query_text(query)	Convenience wrapper returning list of strings
item_count	Property returning number of stored items
get_all_items()	Return all non-expired memory items

Requirements

• Python >= 3.11
• mcal-ai >= 0.2.0
• autogen-core >= 0.4.0 (optional — gracefully degrades if not installed)
• autogen-agentchat >= 0.4.0 (optional)

License

MIT License