klyd

░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░  ░░░░  ░░░      ░░░        ░░░      ░░░  ░░░░  ░░  ░░░░░░░░  ░░░░░░░░  ░░░░  ░
▒  ▒  ▒  ▒▒  ▒▒▒▒  ▒▒▒▒▒  ▒▒▒▒▒  ▒▒▒▒  ▒▒  ▒▒▒▒  ▒▒  ▒▒▒▒▒▒▒▒  ▒▒▒▒▒▒▒▒   ▒▒   ▒
▓        ▓▓  ▓▓▓▓  ▓▓▓▓▓  ▓▓▓▓▓  ▓▓▓▓▓▓▓▓        ▓▓  ▓▓▓▓▓▓▓▓  ▓▓▓▓▓▓▓▓        ▓
█   ██   ██        █████  █████  ████  ██  ████  ██  ████████  ████████  █  █  █
█  ████  ██  ████  █████  ██████      ███  ████  ██        ██        ██  ████  █
████████████████████████████████████████████████████████████████████████████████

Persistent architectural memory for autonomous coding agents.

Current status

Early design & specification phase. Subsystem boundary defined; codebase scaffolded.

Overview

Large Language Models (LLMs) tend to lose architectural context and make regression decisions over long coding sessions. klyd solves this by acting as a persistent architectural memory layer. It extracts architectural decisions from git history/diffs, stores them locally, and retrieves relevant constraints at write time to keep autonomous agents aligned with project guidelines.

klyd remembers. The kernel enforces.

Core Features

• Architectural Decision Extraction: Auto-extracts design constraints and guidelines from repository changes and commit history.
• Local Memory Store: Uses a lightweight local SQLite database to persist constraints offline.
• Context Retrieval: Queries relevant architectural constraints based on the files currently being edited.
• Agent Prompt Injection: Injects retrieved guidelines into agent execution loops to guide code generation.

Relationship to Kernel

• klyd (Memory): Tracks and provides the historical context and architectural rules.
• Kernel (Enforcement): Deteministically evaluates code changes at the write-path to ensure those rules are not violated.

Non-goals

• Write-path interception: klyd does not block files or handle save hooks; that is the sole responsibility of the Kernel.
• Code Generation: klyd does not generate or rewrite source code.

Installation

pip install -e .

Usage

klyd --help
python -m klyd --help

Links

• WatchLLM Organization
• Central Documentation
• watchllm.dev