Context-Infrastructure: Persistent Memory Blueprint for AI Agents

Source: https://github.com/grapeot/context-infrastructure Author: grapeot Date: 2025

Summary

Year-long running reference implementation of persistent memory and context system for AI coding agents. Three-layer architecture: display layer (43 decision axioms, 25+ reusable skills), reusable layer (SOUL.md, USER.md templates), non-reusable layer (personal behavioral data). Goal: move beyond AI that “only speaks correct platitudes” to genuinely personalized context.

Key Claims

  • Problem: AI produces “correct but generic responses” without personalized context
  • Layer 1 (Display): 43 decision axioms + 25+ reusable skills — accumulated over a year
  • Layer 2 (Reusable): SOUL.md (identity), USER.md (preferences), communication style, memory code
  • Layer 3 (Non-reusable): Domain insights; must accumulate yourself; no shortcuts
  • Immediate value: fill in USER.md for instant personalization
  • True value: build your own behavioral dataset over time

Connection to Other Sources

Related to Karpathy LLM Wiki Pattern (same wiki-first approach) and Context Engineering. Also connects to Claude Code Spec Workflow.

Concepts