AbstractLLM v1.0.3

[1.0.3] - 2025-09-14

Added

• Global ALMA Command: Added alma console script that provides global access to the intelligent agent
• CLI Module: New abstractllm.cli module that integrates all SOTA features from alma-simple.py
• Universal Agent Access: Users can now run alma from anywhere after installing AbstractLLM
• Full Feature Integration: The global command includes hierarchical memory, ReAct reasoning, knowledge graphs, and tool support

Changed

• Package Distribution: Enhanced package to include console script entry point
• User Experience: Simplified access to the intelligent agent capabilities without needing to clone the repository

Fixed

• Tool Call Parsing: Enhanced JSON parsing robustness for LLM-generated tool calls with unescaped newlines
• Write File Tool: Fixed tool call parsing when content contains literal newlines or special characters

Installation

After upgrading to v1.0.3, users can install and use the global command:

pip install abstractllm==1.0.3
alma --help
alma --prompt "Hello, I'm testing the global command"
alma  # Interactive mode with memory and reasoning

AbstractLLM v1.0.2

[1.0.2] - 2025-09-14

Fixed

• OpenAI Provider Response Format: Fixed OpenAI provider to return proper GenerateResponse objects instead of raw strings, ensuring consistency with other providers and proper response structure

[1.0.1] - 2025-09-14

Fixed

• MLX Dependencies in [all] Extra: Fixed issue where pip install "abstractllm[all]" did not include MLX dependencies
• User Experience: Users can now install all provider support including MLX using the [all] extra
• Platform Compatibility: Added documentation clarifying that MLX dependencies are Apple Silicon specific

Changed

• Installation Documentation: Updated README to clarify that [all] extra now includes MLX dependencies
• Platform Notes: Added note about MLX platform compatibility

# AbstractLLM v1.0.0 - Major Release 🚀

AbstractLLM v1.0.0 - Major Release 🚀

Overview

This is a major release that transforms AbstractLLM from a basic LLM interface into a comprehensive agentic framework with memory, reasoning, and enhanced tool capabilities.

🎯 Key Highlights

Hierarchical Memory System (Alpha)

• Three-tier memory architecture: Working, Episodic, and Semantic memory
• Cross-session persistence with knowledge consolidation
• Automatic fact extraction and knowledge graph integration

ReAct Reasoning Cycles (Alpha)

• Complete reasoning cycles with scratchpad traces
• Context-aware retrieval for enhanced responses
• Bidirectional linking between memory components

Enhanced Tool System

• Pydantic validation and retry logic
• Universal compatibility across all providers
• Architecture-aware tool handling

Structured Response System

• JSON/YAML formatting with validation
• Works across all 5 providers (OpenAI, Anthropic, Ollama, HuggingFace, MLX)
• Retry strategies with error recovery

🔧 Technical Improvements

Provider Enhancements

• OpenAI: Manual improvements for better tool support
• Universal Tool Handler: Adapts to model capabilities
• Enhanced Architecture Detection: Better model optimization
• Cross-Provider Compatibility: Memory and reasoning work everywhere

Agent Development

• ALMA-Simple Agent: Complete example with memory and tools
• Enhanced Session Management: Persistent conversations
• Tool Integration: Native and prompted modes

📚 New Components

New Files Added

• abstractllm/memory.py - Hierarchical memory system (1860+ lines)
• abstractllm/retry_strategies.py - Advanced retry strategies
• abstractllm/scratchpad_manager.py - ReAct reasoning management
• abstractllm/structured_response.py - Universal structured responses
• abstractllm/tools/enhanced_core.py - Enhanced tool definitions
• Plus many more utilities and enhancements

🚨 Breaking Changes

While the core API remains backward compatible, this release includes:

• Internal architecture restructuring for better maintainability
• New alpha features for memory and reasoning
• Enhanced tool system (additive, not breaking)
• Updated provider implementations

📖 Migration Guide

For Existing Users

# Existing code continues to work
llm = create_llm("openai", model="gpt-4o-mini")
response = llm.generate("Hello world")

New Memory Features (Alpha)

# Enable memory and reasoning
session = create_session(
    "anthropic",
    enable_memory=True,              # Hierarchical memory
    memory_config={
        'working_memory_size': 10,
        'consolidation_threshold': 5
    }
)

# Use memory context and reasoning
response = session.generate(
    "Analyze the project",
    use_memory_context=True,         # Alpha feature
    create_react_cycle=True          # Alpha feature
)

⚠️ Important Notes

• Memory and agency features are in alpha testing
• OpenAI support achieved through manual provider improvements
• Core features are stable and production-ready
• Alpha features are experimental but functional

🔗 Links

• PyPI: https://pypi.org/project/abstractllm/1.0.0/
• Documentation: See updated README.md and docs/ folder
• Examples: Check out alma-simple.py for agent examples

🙏 Acknowledgments

This release represents a significant evolution in AbstractLLM's capabilities, bringing advanced agentic features while maintaining the simplicity and reliability that users expect. This work would not have been possible without the various Open Source communities, libraries and supports.

---

Full changelog: See CHANGELOG.md for detailed technical changes and migration information.