Adam Framework

AI Amnesia.
Solved.

Every time you start a new session, your AI forgets everything. This framework fixes that — with 5 layers of persistent memory and coherence, a neural graph that grows nightly, and 353 sessions of production proof.

View on GitHub Install in 30 min →
353
Sessions
16,200
Neurons
8mo
Production
✦ 353 sessions✦ 6,619 message turns✦ 16,200 neurons✦ 47,874 synapses✦ 4 model migrations survived✦ 1 nuclear reset survived✦ identity preserved through all of it✦ built by one non-coder✦ runs on consumer hardware✦ model-agnostic✦ MIT license ✦ 353 sessions✦ 6,619 message turns✦ 16,200 neurons✦ 47,874 synapses✦ 4 model migrations survived✦ 1 nuclear reset survived✦ identity preserved through all of it✦ built by one non-coder✦ runs on consumer hardware✦ model-agnostic✦ MIT license
The Problem

Your AI wakes up blank. Every. Single. Time.

You've explained your projects, your people, your preferences — dozens of times. The AI is powerful inside a single conversation and completely useless as a long-term collaborator. That's AI Amnesia. The Adam Framework is the architectural fix.

[SESSION 001] Context loaded. Adam knows who you are. [SESSION 047] Neural graph: 3,200 neurons active. [SESSION 119] Reconcile complete. Memory compounds. [SESSION 119] Coherence check: exit 0. Session coherent. [REBUILD] System wiped. Vault files survived. [SESSION 120] Identity intact. Picking up where we left off. [SESSION 353] 16,200 neurons. 47,874 synapses. Still running.   // The AI doesn't forget anymore.
How It Works

Five layers.
One continuous memory.

Each layer solves a specific failure mode that appeared in production. Together they close every gap between session end and session start — and keep the AI coherent through the session itself.

01
Vault Injection
Identity files compiled by SENTINEL before every boot. Your AI knows who it is, what's active, and today's date before it says a single word. Deterministic. No inference. Just files.
02
Session Memory Search
Hybrid vector + text search across all prior sessions and Vault content. 70% semantic, 30% exact match. The AI retrieves what's relevant mid-conversation without loading everything upfront.
03
Neural Graph
A local SQLite knowledge graph with spreading activation, Hebbian learning, and temporal decay. Concepts link to concepts. Context propagates. 16,200 neurons built from 353 sessions — and growing every night.
04
Nightly Reconciliation
When the context window approaches its limit, the AI writes its own memory flush before truncation. Nothing is lost. The sleep cycle merges daily logs into core memory via Gemini. Memory compounds while you sleep.
05
Coherence Monitor
SENTINEL runs a scratchpad dropout detector every 5 minutes during active sessions. When the AI stops using its reasoning loop — a real, binary signal for within-session drift — a targeted re-anchor fires into BOOT_CONTEXT.md before damage is done. Within-session coherence degradation: solved.
The Proof

Not a demo.
A production record.

Every number here comes from a live system running real business operations — built over 8 months, in production since day one of Adam’s creation. Not a benchmark. Not a controlled test.

353
Total sessions across 8 months of production use
6,619
Message turns — the full conversational record
16,200
Neural graph neurons — live, updated nightly
47,874
Neural graph synapses — growing with every reconcile
4
Model migrations survived with no memory loss
1
Complete nuclear reset survived — identity intact

“When the system was completely wiped and rebuilt, the AI came back online with full continuity — because the Vault files survived. Same base model. Same files. Same AI.”

— docs/PROOF.md · The nuclear reset validation, February 2026
The Story

One non-coder.
Eight months.
One nuclear reset.

This wasn't built in a lab. It was built by someone running a small business in Miami who needed their AI to remember things — and kept fixing it until it did.

Late January 2026
ClawdBot — the beginning
First version. No persistence. The AI reset completely between every session. The problem was obvious within weeks.
Early February 2026
Layer 1 emerges
Write everything down in files, make the AI read them. SOUL.md and CORE_MEMORY.md created. Session coherence improved immediately.
January–February 2026
The compaction problem, then the fix
Important context was getting pushed out as sessions grew. Fix: the AI writes its own memory before truncation. Layer 4 born.
February 14–16, 2026
The nuclear reset
System wiped and rebuilt in 10 hours. The AI came back online knowing every active project — because the Vault files survived. Architecture validated.
February 25–28, 2026
Full stack complete
Neural graph integrated. SENTINEL hardened. All four memory layers running simultaneously.
March 3, 2026
Public release
353 sessions. 12,393 neurons. 8 months of continuous operation. Templated, documented, and shipped.
March 5, 2026
Layer 5: Within-Session Coherence Degradation Solved
Scratchpad dropout identified as a binary, production-validated signal for within-session AI drift. coherence_monitor.py ships with 33/33 tests passing against live session data. SENTINEL checks every 5 minutes. Re-anchor fires into BOOT_CONTEXT.md when drift is detected. Problem Two: solved.
Read the full story →
Get Started

30–60 minutes.
Two paths.

You have OpenClaw already installed and running. From there, pick your path.

PATH 01 — YOU DO IT
Human install guide
Step-by-step for humans. Four phases: identity files, neural memory, Session 000 history seeding, sleep cycle. No assumptions made.
Read SETUP_HUMAN.md →
PATH 02 — YOUR AI DOES IT
Agent install guide
Give this doc to your existing AI assistant. It asks you 8 questions and handles the install. Written for AI agents: terse, command-first, explicit success conditions.
Read SETUP_AI.md →
# Seed your neural graph with your full conversation history
python tools\legacy_importer.py --source export.zip --vault-path C:\MyVault --user-name YourName

# Ingest into the neural graph (~56 min for 740 facts)
.\tools\ingest_triples.ps1 -VaultPath C:\MyVault