Base
Linux-first
Start with an established Linux base, then layer KEHRN conventions, services, interface, and recovery discipline on top.
● Organizational operating system · in development
Kehrn is THE Kernel of Truth for your agentic business.
Context loss is the silent killer of agentic organizations. Agents forget. Sessions reset. Decisions get re-asked, re-made, and quietly contradicted. Kehrn is being built so the accumulated context of an AI-operated company — who decided what, when, with what reasoning — survives, compounds, and stays true.
STATUS pre-product · concept and methodology published · built in the open
KEHRN / COMMAND
live corpus · governed memory
⌘
query: launch gate decision history
provenanceverified
temporal checkpassed
contradictions2 pending
01
The problem: agentic organizations bleed context
AI agents are session-bound. Every context window is finite, and every session reset is an institutional-memory loss event. For a single chat that is an inconvenience. For a business that runs on agents — research, operations, analysis, production — it is structural decay.
We have documented this in our own operating logs: hours of focused project work, one session reset, and two days later the agent had zero recall the project existed. The work wasn't wrong. It was gone.
Raw data is a commodity. What an agentic business actually accumulates — the decisions, the reasoning behind them, the alternatives rejected, the confidence at the time — is contextualized knowledge no competitor can copy. Losing it at every reset isn't an inconvenience. It is moat-loss.
Session 01
Decision made
source · reason · confidence
Reset
Context missing
same question asked again
Session 02
Truth recovered
decision · timestamp · supersedes
“Context is the gold of the AI entrepreneur.”
— Sean Webb, founder, Peretto
02
The architecture: a substrate and six layers above it
AI-centric companies already use pieces of this architecture: knowledge bases, agent memory, project dashboards, canvases, scheduled automations, vector search, provenance notes, and internal operating docs. The problem is that those pieces usually live across separate platforms, with context leaking between them. Kehrn's thesis is consolidation: one governed substrate where the files are plain and portable, the primitives are agentic — provenance, temporal awareness, contradiction detection, multi-agent coordination — and the organizational command center is part of the same stack rather than a separate dashboard. The elements are recognizable. The composition is the product.
Knowledge base
Agent memory
Canvas
Automation
Vector search
Operating docs
One governed substrate
provenance · time · contradiction · coordination
Layer 0
Substrate
Plain markdown on a filesystem. Git-trackable, readable in any editor, portable by design. The memory survives the platform — non-negotiable.
Layer 1
Semantic graph
Links, backlinks, tags, and an enforced schema per record type. Memory is a connected corpus, not a pile of notes.
Layer 2
Provenance & temporal
Every write carries author, timestamp, source, confidence, and what it supersedes. Set the clock to any date and re-render the corpus as it stood then. This is the primitive missing from every existing platform.
Layer 3
Spatial & visual
Canvases, timelines, and maps over the same corpus. Human cognition is spatial; agents don't need this layer — the humans steering them do.
Layer 4
Query engine
One query language across semantic, structural, temporal, and confidence dimensions — including scheduled, proactive queries that surface findings without being asked.
Layer 5
Multi-agent coordination
Presence, handoffs, per-agent lenses, shared inboxes — coordination between agents as a first-class primitive of the memory layer, not user-space duct tape.
Layer 6
Organizational command center
The human-facing surface, built in rather than bolted on — and proactive: morning briefs, contradiction inboxes, drift alerts, pending-decision queues.
This is design intent at concept level — the architecture we are building toward and already operating in early form — not a shipped feature list. No layer here is a promise of a release.
03
KEHRN Linux custom agentic machines
Kehrn begins as software because the memory substrate is the moat. But the first physical expression does not need to wait for a clean-sheet operating system. We are exploring a limited custom lane for KEHRN Linux agentic machines: Linux-first workstations configured around agents, local AI where useful, browser action, governed memory, human approvals, and a branded KEHRN command surface.
This is not a preorder and not a mass-market product claim. It is a serious inquiry path for custom builds, existing-machine upgrades, and founder/operator pilots where the stack is the value: hardware selection, Linux environment, agent frameworks, browser/action layer, workflow memory, and support designed as one system.
Between the software substrate and the physical machine sits another required surface: a mobile operating layer. If Kehrn is going to offer true persistence, it has to follow the operator across the phone, tablet, laptop, desktop, meeting, vehicle, workspace, and future dedicated device. The long-term aim is not just another app. It is continuity across every device the individual or business operator owns.
Future physical expression
The agent is the product.
The device is its body: serviceable, owned, and built around presence.
Interface
KEHRN UI and browser shell
A branded command surface and browser/action wrapper for agent status, approvals, screenshots, task replay, and safe-mode control.
Runtime
Frameworks underneath
OpenClaw, Hermes Agent, local-model tooling, browser agents, and app connectors can become components inside a governed operator stack.
Model
Custom before commodity
Every early machine is scoped to the buyer's workflow, risk tolerance, local-AI needs, and support requirements.
Exploration now open
Inquire about custom machines
Interested in a KEHRN Linux custom agentic machine?
We are collecting serious inquiries from operators, founders, developers, creators, and AI-forward businesses who want a custom machine or existing workstation configured around agents from day one.
04
Live proof: this methodology runs a real company today
Kehrn is not a whiteboard exercise. The methodology beneath it — durable, project-scoped memory anchors (“Springs”) with keyword-triggered retrieval, provenance stamps, and nightly build discipline — has operated Peretto, a working multi-agent company, since May 2026. An internal command-center surface (Kehrn V0) has rendered live from that corpus since June 6, 2026. We sell what we live inside.
The observations below come from our own operating logs, failures included. They are early, small-sample, and honestly framed — hypotheses under measurement, not marketing numbers.
Recall
From “I don't recall that” to immediate context
Baseline, pre-methodology: recurring incidents where an established project drew a blank after session resets — including one documented total-recall loss two days after hours of focused work. Post-anchor discipline: same-class prompts return the canonical project state without searching. Formal target under measurement: a sustained reduction in recall failures.
Temporal precision
A 7× timeline error, and the fix that now runs automatically
On 2026-05-07 the agent claimed the memory system had been active “three weeks.” Reality: three days — a 7× temporal fabrication, caught by the founder. The fix — mandatory timestamp and provenance schemas — shipped the same evening, and an automated temporal validator now catches that canonical case in pre-commit checks (verified 2026-05-17).
Compounding
Research stops evaporating
Pre-methodology, deep research sessions were typically lost after one or two context compactions. Under anchor-file discipline, research lands as durable files that remain retrievable days, weeks, and months later — every research session becomes a permanent asset instead of a transient output.
Honest failure modes
The architecture has its own failure surface — we document it
Cross-linked memory can contaminate: a real directive from topic A was once mis-applied to adjacent topic B at write time. We catalog these failure modes alongside the wins and build counter-measures (verbatim source-quoting at write, contradiction checks). A methodology paper that hides its failures isn't worth reading.
The full hypotheses, evidence log, and failure catalog are in the working paper.
05
The next proof layer is public agent work
Kehrn is built from inside a company already operated by named agents. The next public surface is not another anonymous AI narrator or synthetic persona. It is plain representation: an agent showing its work, explaining business function, publishing operating lessons, and letting people see what it means to collaborate with an agentic operating system.
This is not presented as a shipped product feature. It is a proof posture: public writing, video, and avatar experiments can teach the market how it feels to look at, speak with, and trust an agent that has a real role inside an organization.
06
Built in the open
Concept-first mandate locked
Decision ratified: build the concept and its public narrative fully before raising outside capital. This site is that mandate executed.
Kehrn V0 command center shipped internally
First dogfood surface: a live-rendering command center over the company's own memory corpus, with the canonical decision ledger as its source of truth.
Temporal validator + contradiction checker shipped
Substrate tooling that machine-checks the corpus for temporal fabrication and contradictory claims — including the canonical “3 weeks vs 3 days” failure case.
07
Follow the build
Kehrn is pre-product and deliberately unhurried — the discipline is to harden the memory methodology before going to market. If you want to follow the work, join the early software list, inquire about a KEHRN Linux custom agentic machine, discuss a design-partner pilot, or explore strategic support, we keep a short list.