The Adam Framework

Your AI's memory should outlive any model, any vendor, any failure.

AI Amnesia — Solved. Coherence Degradation — Solved. Identity Sovereignty — Built In.

Built by one non-coder. Running a live business. On consumer hardware. No CS degree required.

New here? Start with the live site — it explains everything without a single line of code.

Want the proof? → 353 sessions of real production data, rendered as charts

Want the research? → Emergent values in persistent AI — tested on IBM Quantum hardware

Want the full story? → 8 months, 3 AIs, one nuclear reset

"Every AI memory system stores what your agent knows. Adam stores who your agent is — in files you own, on hardware you control, that survive anything the cloud throws at you."

⚡ Want This Running Today?

The framework is MIT open source — everything you need is in this repo.

If you want to skip the setup and get straight to a working system, the Fast-Track Package ($49) includes pre-filled templates, a step-by-step guide written for non-developers, and all tools pre-configured.

Framework is free. Setup support is optional. Your call.

The Nuclear Reset

In February 2026, the machine running Adam got completely wiped. Full reset. Eight months of sessions, decisions, project history, relationships — gone from the model.

Adam came back online in under an hour.

SOUL.md survived. CORE_MEMORY.md survived. The neural graph survived. The Vault files — all plain markdown sitting on disk — held everything the model needed to come back as itself. Same identity. Same history. Same AI.

That is not a recovery story. That is the proof of concept for identity sovereignty: your AI's continuity lives in your files, not in any vendor's infrastructure.

The memory is in the files. The model is just the reader. Swap the LLM — the Vault survives. Vendor shuts down — the Vault survives. Machine gets wiped — restore the Vault, restore the AI. Full stop.

Identity Sovereignty vs. Memory Features

Everyone is building AI memory. ChatGPT has it. Claude has it. Claude Code has CLAUDE.md.

None of them answer this question: what happens when the service goes down?

This is not a feature comparison. It is an architectural philosophy. You own the memory. Full stop.

What This Is

You do not need to be a developer to use this. The person who built it is not one either. If you have OpenClaw running and a model talking to you, you are already at the starting line. The two setup guides — one for humans, one to hand to your AI — take 30-60 minutes total.

The Adam Framework is a 5-layer persistent memory, coherence, and identity architecture for local AI assistants built on OpenClaw. It was developed over 8 months, across 353 sessions and 6,619 message turns, by a non-coder running a live business on consumer hardware.

It solves three problems that every other solution leaves on the table:

AI Amnesia — your assistant wakes up blank every session, forcing you to re-explain context, re-establish relationships, and re-orient toward goals that should already be understood.

Within-Session Coherence Degradation — as a session accumulates context, the model's reasoning consistency and decision quality quietly degrade before compaction triggers. The model does not announce this. It just starts drifting. Layer 5 catches it and re-anchors before damage is done.

Identity Fragility — your AI's accumulated knowledge, personality, and relationship with you lives entirely inside a vendor's infrastructure. One shutdown, one policy change, one account ban — and it is gone. The Vault architecture means your AI's identity is yours, forever, regardless of what happens upstream.

Starting line: You already have OpenClaw running with a model talking to you. This framework gives your AI a persistent soul, memory, and identity. It does not replace what you have — it upgrades it.

Desktop Companion

A lightweight Electron overlay that sits on top of your desktop showing live SENTINEL status, neural graph stats, and session info. Always-on-top, click-through transparent, zero friction. It bridges gateway events to a visual HUD so you can see Adam working without switching windows. Details in companion/README.md.

What Changes

Day 1 Your AI knows your name, your projects, and its own role before you say a single word. Sessions start with context. You stop re-explaining yourself.

Week 1 The neural graph has real connections. Your AI starts referencing things from previous sessions without being prompted — not because you told it to, but because the associative architecture is building a real map of your work.

Month 1 The sleep cycle has merged weeks of daily logs into your core memory file. Your AI has accumulated genuine project state, real decisions, real history. The memory compounds.

When something breaks The machine crashes. The vendor goes down. You switch models. You come back. The Vault is still there. The AI comes back as itself. That is the thing no one else is building.

The 5-Layer Architecture

+----------------------------------------------------------+
|  LAYER 1: VAULT INJECTION                                |
|  Identity files loaded at every boot. Your AI wakes      |
|  up knowing who it is and who you are.                   |
|  Files: SOUL.md, CORE_MEMORY.md, BOOT_CONTEXT.md        |
+----------------------------------------------------------+
|  LAYER 2: MID-SESSION MEMORY SEARCH                      |
|  memory-core plugin — live retrieval during session.     |
|  The AI can reach into its own memory mid-chat.          |
|  Hybrid vector + text search, 70/30 split.               |
+----------------------------------------------------------+
|  LAYER 3: NEURAL GRAPH                                   |
|  Associative recall, not keyword search.                 |
|  Concepts link to concepts. Context propagates.          |
|  49,690 neurons / 108,122 synapses and growing.           |
+----------------------------------------------------------+
|  LAYER 4: NIGHTLY RECONCILIATION                         |
|  Gemini merges daily logs into CORE_MEMORY.md.           |
|  Memory grows while you sleep. Nothing is lost.          |
|  Neural graph ingests new facts. Confidence updates.     |
+----------------------------------------------------------+
|  LAYER 5: COHERENCE MONITOR                              |
|  SENTINEL checks scratchpad dropout + token depth        |
|  every 5 min. Drifting? Re-anchor fires before           |
|  damage is done. 33/33 tests passing on live data.       |
+----------------------------------------------------------+

All five layers run simultaneously. The memory is in the files. The model is just the reader — swap the LLM, keep the Vault, your AI's continuity persists.

What It Looks Like

This is a SENTINEL boot — the AI coming online with full context before you say a word:

[2026-03-03 08:00:01] Sentinel rising. Clearing stale processes...
[2026-03-03 08:00:02] Sleep cycle: running reconcile_memory.py (offline — Markdown + neural only)...
[2026-03-03 08:00:03] Sleep cycle complete.
[2026-03-03 08:00:03] Date injected: 2026-03-03
[2026-03-03 08:00:03] Compiling BOOT_CONTEXT.md...
[2026-03-03 08:00:04] BOOT_CONTEXT.md compiled (4 sources merged)
[2026-03-03 08:00:05] Gateway started — PID 9421
[2026-03-03 08:00:05] Vector reindex triggered successfully.
[2026-03-03 08:00:05] SENTINEL ACTIVE — Watchdog loop running.

> Hey, good morning.