SkyLAM

SkyLAM is a local autonomous language-model framework designed for persistent AI-assisted development and system automation.

It combines configurable language models, semantic memory, agentic tool use, and a persistent execution environment into a single runtime.

Features

• Two-Stage LLM Pipeline Uses separate decoder and chat models, allowing SkyLAM to dynamically determine generation parameters before producing a response.

• Persistent Semantic Memory Stores conversation history and embeddings, enabling SkyLAM to retrieve relevant information from previous interactions while maintaining recent context.

• Context Management Automatically balances recent conversation history with semantically relevant memories to operate within model context limits.

• Multimodal Model Support Supports compatible vision-language models in addition to standard causal language models.

• Configurable Model Runtime Supports Hugging Face models, optional PEFT adapters, automatic device selection, memory-aware model loading, and configurable RoPE scaling.

• Agentic Tool System Allows the model to interact with its environment through structured tool calls, including filesystem operations, directory management, command execution, and system inspection.

• Python Execution Provides the agent with controlled access to Python execution for automation and development tasks.

• Persistent Agent Lifecycle SkyLAM can run continuously in the background, recover its previous state, process administrative commands, and restart itself when necessary.

• Administrative Interface Provides CLI controls for launching, configuring, monitoring, messaging, restarting, and shutting down the runtime.

• Session Management Supports multiple persistent conversation sessions with independent histories and embeddings.

• KPCore / KDPH Package System Includes an encrypted package format for distributing, building, extracting, and managing SkyLAM-like packages (KPs) and dependencies.

• GitHub Package Distribution KDPH can publish and retrieve encrypted packages through GitHub-based package repositories.

To Install

KP Workflow

git clone https://github.com/Knexyce-Co/skylam-agent.git && cd skylam-agent && ./ready.py

PIP Install (No KP logic executed on the user end.)

pip install skylam

Architecture

SkyLAM is organized around several core components described below:

SkyLAM
├── Model Runtime
│   ├── Decoder Model
│   ├── Chat Model
│   └── Embedding Model
│
├── Agent Runtime
│   ├── Memory
│   ├── Context Management
│   ├── Tool Execution
│   └── Python Execution
│
├── Administration
│   ├── Configuration
│   ├── Process Management
│   ├── Sessions
│   └── Logging
│
└── KPCore
    ├── Package Creation
    ├── Encryption
    ├── Dependencies
    └── Distribution

Quick Start

Configure the required models:

python -m skylam -c decoder_engine=<Stage I model here.>
python -m skylam -c chat_engine=<Stage II model here.>
python -m skylam -c embed_engine=<Embedding model here.>

Then launch SkyLAM:

python -m skylam --launch

Additional runtime controls are available through the CLI:

python -m skylam --configure key=value
python -m skylam --send "message"
python -m skylam --view --execute
python -m skylam --quit

Configuration

SkyLAM supports configuration for:

• Decoder model.
• Chat model.
• Embedding model.
• Active session.
• PEFT adapter.
• RoPE scaling.

Configuration is stored locally and can be modified through the command-line interface.

KPCore

SkyLAM includes KPCore, a package and distribution system built around the Knexyce Data Package Handler (KDPH).

KPCore provides:

• Encrypted package creation and extraction.
• Package metadata.
• Dependency management.
• Build hooks.
• Local package installation.
• GitHub-based package distribution.

This allows SkyLAM projects and extensions to be packaged and distributed as self-contained encrypted packages.

Project Structure

skylam/
├── boot/        Dependency and startup management.
├── config/      Runtime configurations. (Placeholder null values by default.)
├── console/     Agent and administrator interfaces.
├── core/        Model runtime and memory.
├── kpcore/      Package system.
├── lock/        Runtime state and coordination.
├── oslayer/     Platform-specific functionality.
├── util/        Runtime utilities.
└── ...          Additional software or data may be created by SkyLAM.

Requirements

• Python 3.10+
• PyTorch
• Hugging Face Transformers
• Sentence Transformers
• Accelerate
• PEFT
• NumPy
• Additional dependencies are handled by SkyLAM's bootstrap system.

SkyLAM is intended to provide a persistent, extensible, and locally controlled AI runtime capable of reasoning, remembering, executing tools, modifying its environment, and managing its own lifecycle.