By Jereme Strange · also on Strange Intel
On March 31st, 2026, Anthropic accidentally shipped a source map file inside their npm package for Claude Code. Within hours, 512,000 lines of TypeScript were reconstructed and published to GitHub. It became the fastest repo in history to hit 100K stars.
Inside that code, the community found something that caught my attention: an unreleased feature called **KAIROS**.
KAIROS is an autonomous background agent. It runs without user input. It has background daemons, memory distillation that consolidates knowledge overnight, cron-scheduled tasks, and webhook-driven triggers that let it respond to events automatically. It was fully built but gated behind a compile flag. Never shipped.
Here's the thing. I already built the same architecture. Independently. Months before anyone knew KAIROS existed.
And I can prove it.
---
## The Receipts
The Adam Framework repo has been public on GitHub since March 3rd, 2026. Every commit is timestamped and immutable:
- **March 3**: Repo created. SOUL.md, CORE_MEMORY.md, vault architecture.
- **March 8**: SENTINEL watchdog daemon committed. Background process that monitors health, auto-restarts on failure, runs coherence checks every 5 minutes. This is the KAIROS daemon pattern.
- **March 8**: Sleep cycle reconciliation. Nightly memory consolidation that merges daily logs into core memory using an LLM. This is the KAIROS memory distillation pattern.
- **March 10**: Neural graph integration. Associative memory with spreading activation, Hebbian learning, temporal decay. 16,200+ neurons, 47,000+ synapses.
- **March 14**: AGENTS.md and SWARM architecture documented. Multi-agent coordination through shared vault files. No message passing between agents — the filesystem is the queue.
- **March 19**: Coherence monitor with 33-test suite. Detects within-session identity drift before compaction triggers. Fires re-anchor injection when threshold exceeded.
The leak happened on March 31st. Every one of these commits predates it by weeks.
I'm not claiming I knew about KAIROS. I didn't. I'm claiming something more interesting: **the problem demands this architecture**. If you spend enough time trying to make an AI agent actually work in production — not in a demo, not in a tutorial, in a real business — you arrive at the same design.
Background daemon. Memory consolidation. Coherence monitoring. Event-driven triggers. These aren't features. They're requirements.
---
## What KAIROS Has vs. What Adam Has
| Capability | KAIROS (Anthropic, unreleased) | Adam Framework (public since March 3) |
|---|---|---|
| Background daemon | Yes — runs without user prompts | Yes — SENTINEL watchdog, auto-restart, health checks |
| Memory distillation | Yes — autoDream nightly consolidation | Yes — reconcile_memory.py, Gemini-powered merge of daily logs |
| Cron scheduling | Yes — cron-scheduled refreshes | Yes — SENTINEL runs on Windows Task Scheduler |
| Webhook triggers | Yes — GitHub webhook subscriptions | Yes — Event bus with HTTP webhook receiver (shipped April 3) |
| Multi-agent coordination | Yes — internal multi-agent system | Yes — SWARM file-based task orchestration with atomic locks |
| Coherence monitoring | Not documented in leak | Yes — Layer 5, 33-test suite, re-anchor injection |
| Identity persistence | Not documented in leak | Yes — 5-layer vault architecture, survived full system wipe |
| Provider agnostic | Locked to Claude | Yes — works with Gemini, OpenAI, Anthropic, Ollama, any compatible endpoint |
The Adam Framework is open source. MIT licensed. `pip install adam-framework`.
---
## Why This Matters
The Claude Code leak revealed something important: **the model is not the innovation. The harness is.**
Everyone is focused on which model is smarter. GPT-5 vs Claude 4 vs Gemini Ultra. That's the wrong question. The right question is: what wraps around the model to make it actually useful?
The answer — and this is what both KAIROS and Adam Framework arrived at independently — is an orchestration layer. A harness. Something that manages memory, maintains identity, coordinates agents, handles events, and keeps the whole thing coherent over time.
Anthropic built this internally and hid it behind a compile flag. I built it independently and shipped it open source.
The entire "harness engineering" movement that emerged after the leak — the blog posts, the frameworks, the tutorials — is converging on the same architecture we've been running in production for months.
---
## The Numbers
This isn't theoretical. Adam Framework runs a live business:
- **353 sessions** of production use
- **6,619 message turns** processed
- **18,844 neurons / 51,815 synapses** in the neural graph (growing nightly)
- **4 model migrations** survived without data loss
- **1 complete system wipe** survived — full nuclear reset, recovered in under an hour
- **33/33 coherence monitor tests** passing against live session data
The memory is in the files. The model is just the reader. Swap the LLM, keep the vault, continuity persists.
---
## What I'm Building Next
The KAIROS leak confirmed we're on the right track. Now we're pushing further:
- **SWARM runtime** — file-based multi-agent task orchestration is now real code, not just a design doc. Atomic locks, agent registry, priority queues, fail/retry lifecycle.
- **Event-driven triggers** — HTTP webhook receiver so external systems (GitHub, email, social) can trigger agent actions automatically.
- **Provider abstraction** — one interface for Gemini, OpenAI, Anthropic, Ollama, or any compatible endpoint. Switch models without touching your memory or configuration.
- **CLI tool** — `adam init`, `adam start`, `adam status`. The framework is now pip installable.
All of this shipped on April 3rd. The commit is public.
---
## How to Follow This Work
I'm documenting everything I build in real-time. No scripts. No theory. Actual production systems, actual business results, actual code.
**GitHub:** [github.com/strangeadvancedmarketing/Adam](https://github.com/strangeadvancedmarketing/Adam)
**Install:** `pip install adam-framework`
---
*Built by Jereme Strange. One non-coder. Consumer hardware. No CS degree. No team. Just AI and a clear vision.*
Want this in your business? Free 30-minute consultation