Context Management with APM

NPM for Agent Context

Just like npm revolutionized JavaScript by enabling package reuse, APM creates an ecosystem for sharing agent context.

Package Composition & Reuse

# Your project inherits team knowledge via apm.yml file in the root
dependencies:
  apm:
    - company/design-system     # UI patterns, brand guidelines
    - company/security-standards # Auth patterns, data handling
    - community/best-practices  # Industry standards

Result: Your project gets all the instructions of above packages applied via dynamically generated Agents.md files using specify apm compile. These files are optimally generated to minimize contextual load for Agents compatible with the Agents.md standard.

Enterprise Scenario: Design team creates accessibility guidelines once → entire organization uses them → agents work consistently across all projects.

Mathematical Context Optimization

The Technical Foundation: APM uses mathematical optimization to solve the context efficiency problem.

Context_Efficiency = Relevant_Instructions / Total_Instructions_Loaded

Why This Matters: When agents work in /styles/ directory, they shouldn't load Python compliance rules. APM's Context Optimization Engine ensures agents get minimal, highly relevant context.

The Algorithm: Constraint satisfaction optimization that finds placement minimizing context pollution while maximizing relevance. Each instruction gets mathematically optimal placement across the project hierarchy.

Quick Start

specify init my-project --use-apm --ai copilot
specify apm install company/design-system  
specify apm compile  # Mathematical optimization generates distributed AGENTS.md files

Universal Agent Compatibility

APM generates distributed AGENTS.md files compatible with the agents.md standard, working with any coding agent (GitHub Copilot, Cursor, Claude, Codex, Aider, etc.).

Authentication Setup (Optional)

export GITHUB_APM_PAT=your_fine_grained_token_here

Only needed for private packages. Public community packages work without authentication.

The Complete Value

Package Ecosystem - Share and compose agent intelligence like code dependencies
Mathematical Optimization - Context Optimization Engine ensures relevance without pollution
Universal Standards - Works with any agent via industry-standard agents.md format
Enterprise Ready - Team knowledge scales across entire organizations

2.5 KiB Raw Blame History