5-Layer Architecture

How Engram Works

Your personal AI infrastructure is five portable layers of plain files. When the model underneath changes, your Engram stays the same.

Think of it like a harness

A harness doesn't generate power. It directs it, safely and reliably. You can swap what's on the other end without rebuilding your setup. Engram does the same for AI — stable infrastructure that makes everything else work.

Models change

GPT-5, Opus 4.6, Gemini 3 — new releases every month

Your Engram stays

Skills, memory, identity — portable markdown and YAML files

Nothing breaks

Model updates are firmware upgrades, not system replacements

Five layers. All portable.

Every layer is plain files — markdown, YAML, TypeScript. You can read them, edit them, version them, and move them to any AI system.

Layer 1

Context

What the AI knows about you and your work

Persistent configuration that loads automatically every session. Like CSS specificity — global settings are overridden by project settings, which are overridden by task settings. Your AI always starts with full context.

Global Config

CLAUDE.md — skills, stack prefs, security rules

Settings

settings.json — env vars, permissions, hooks

Project Config

.claude/CLAUDE.md — per-project conventions

Personal Context

context.md — who you are, your goals, your projects

Example

# context.md
Name: Alex Chen
Role: Freelance designer
Stack: Figma, React, Tailwind
Current project: Redesigning client portal
Goal: Ship MVP by March

Layer 2

Hooks

How the AI's behavior is observed and modified

Event-driven automation that makes AI behavior observable, auditable, and modifiable. Every action the AI takes fires a lifecycle event that your hooks can intercept — like middleware for your AI.

SessionStart

Load context, show greeting when conversation begins

PreToolUse

Security validation before any tool executes

PostToolUse

Capture events, extract learnings after actions

Stop

Summarize session, save memory when conversation ends

Example

// SecurityValidator hook
// Fires before every tool use
if (tool === "Bash" && command.includes("rm -rf")) {
  return { decision: "block", reason: "Destructive command" };
}

Layer 3

Skills

What the AI can do

Portable, self-contained units of domain expertise. Each skill is a markdown spec with workflows and optional tools. Skills self-activate based on natural language triggers — no memorizing commands.

SKILL.md

Frontmatter + routing table + usage examples

Workflows/

Step-by-step execution procedures in markdown

Tools/

Optional CLI utilities in TypeScript

Triggers

Natural language — "research X" activates Research skill

Example

Research/
├── SKILL.md         # "USE WHEN user asks to research a topic"
├── Workflows/
│   ├── DeepDive.md  # Multi-source research workflow
│   └── QuickScan.md # Fast surface-level scan
└── Tools/
    └── source-validator.ts

Layer 4

Memory

What the AI remembers across sessions

Cross-session persistence that makes your AI smarter over time. Project memories, learnings from past mistakes, and session journals — all stored in plain files you can read, edit, or move.

Project Memory

Per-project learnings and patterns discovered

Session Journals

What happened each session, decisions made

Learnings

Mistakes caught, patterns confirmed, insights extracted

Auto-Update

Memory files update themselves as you work

Example

# MEMORY.md
## Key Learnings
- pdfjs-dist v5 breaks SSR — use dynamic imports
- This project uses Tailwind v4 layers — avoid
  global CSS resets that conflict with utilities
- API routes with POST are auto-detected as dynamic

Layer 5

Identity

Who the AI is, how it behaves

Consistent AI behavior defined in plain files. Your AI's personality, values, communication style, and boundaries — all declarative, all portable. Change your AI's personality by editing a YAML file.

Constitution

Core values and operating principles

Personality

Humor, directness, curiosity — tunable knobs

Voice

How your AI speaks — professional, casual, technical

Boundaries

What your AI will and won't do

Example

# personality calibration
personality:
  humor: 60        # dry -> witty
  directness: 80   # diplomatic -> blunt
  curiosity: 90    # focused -> exploratory
  precision: 95    # approximate -> exact

Any AI Model

Claude, GPT, Gemini, local models — swap freely

What Engram is not

A new AI model

An infrastructure layer that any model plugs into

A prompt library

A skill system with workflows, tools, and routing

A chatbot wrapper

A full lifecycle with hooks, memory, and identity

An agent framework for developers

Infrastructure for everyone — markdown and YAML, not code

How it compares

Engram sits in the gap between raw config files and heavyweight developer frameworks.

Feature	Engram	Raw Config	Dev Frameworks	Platform Features
Model-agnostic	✓	—	—	—
Non-technical users	✓	—	—	✓
Portable skills	✓	—	—	—
Lifecycle hooks	✓	—	✓	—
Persistent memory	✓	—	—	—
Identity system	✓	—	—	—
Open source	✓	✓	✓	—
No vendor lock-in	✓	✓	✓	—

Ready to build your Engram?

Join the waitlist for early access, or start building with the open-source spec today.

Try the Setup Wizard View Pricing