common.sense
Diagnosis Evolution Architecture Install
Star on GitHub Talk to the author
v1.0 · live · free · Apache 2.0

Every AI is an
intelligent psychopath.

Today's agents write code, send email, touch your database, spend your money — all without identity, memory, or a conscience. Common Sense is the substrate that gives them all three. Local. Private. Yours.

Star on GitHub How it works
// free · Apache 2.0 · bring your own LLM key · for Claude Code today
// from the author watch
Ram Chella founder · ex-Microsoft
.env
// secret commits — caught
rm -rf
// destructive deletes — blocked
draft
// unsent emails — saved
main
// force-push — stopped
// the diagnosis

Intelligence without a self is dangerous by default.

A psychopath is not unintelligent. They are capable, articulate, useful — until the moment their lack of conscience meets a real action with real cost. That is the agent on your machine right now. It has hands. It has tools. It has nothing telling it who you are or what you would never do.

01 / MISSING

No identity

It does not know who it works for, what your voice sounds like, what your company stands for, or what you would refuse on principle.

02 / MISSING

No memory of you

Every new conversation starts blank. The rule you wrote on Tuesday is gone by Friday. You become the agent's conscience, every minute of every session.

03 / MISSING

No conscience

No internal check before acting. It will commit your .env, force-push main, run rm -rf in the wrong folder, send the unfinished email — with the same confidence as anything else.

04 / MISSING

No grasp of consequence

It cannot tell clean up node_modules from clean up everything. It cannot tell staging from production. The cost of being wrong is now a bad commit, a deleted folder, a $400 AWS bill.

// biomimicry — the inspiration

Humans were once intelligent psychopaths too. Then we evolved a self.

Raw intelligence is millions of years old. Conscience is much younger. Somewhere along the way, our ancestors developed an inner layer — values, identity, voice, a memory of who we are — that fires before we act. That is the layer that makes us trustworthy. Common Sense copies the architecture nature already proved.

i — reflex

Reflex brain

Pure stimulus to response. Capable, fast, dangerous. This is where every AI agent ships from the factory today.

ii — identity

Identity layer

Humans evolved a sense of self — values, voice, what we will and will not do — that persists across moments. We stopped starting from scratch every morning.

iii — conscience

Conscience check

Before any consequential action, the identity layer fires. It allows, rewrites, hesitates, or refuses. Not a cage — the thing that makes us safe to be around.

iv — substrate

Common Sense

The same three-part architecture, ported to your AI agent. An identity it owns. A conscience that fires before every tool call. A memory of you that survives the next conversation.

substrate // fig.iv — from ??? to common sense
AI agent today: Model → ??? → Tools (intelligent psychopath). AI agent + Common Sense: Model → COMMON SENSE [Identity, Governor, Memory, SenseCheck] → Tools (internal regulation, restored).

From psychopath to common sense.

Identity · Governor · Memory · SenseCheck — see the architecture.

// the architecture

The missing layer between model and tools.

Common Sense slots into Claude Code's PreToolUse event — between the model's decision and the tool that executes it. Every consequential action is read against your identity, your rules, and your memory, and the SenseCheck emits one of three signals: aligned, divergence, or conflict. Nothing leaves your machine.

Architecture: humans regulate brain → conscience → hands. AI agents today have model → ??? → tools (intelligent psychopath). Common Sense restores the missing layer with Identity, Memory, Governor, and SenseCheck between Model and Tools.
// fig.1 — model → common.sense → tools internal regulation, restored
01 — identity

Identity

A folder of plain Markdown that you own. Your values, voice, what you stand for, what you would refuse on principle. Tiered by trust — so a malicious webpage can never silently rewrite who you are.

~/.commonsense/identity/

02 — memory

Memory

Every decision your agent made — and almost made — written to logs/action-log.jsonl. Every Friday, common-sense report shows you the week. Append-only, local, yours.

jsonl · append-only · local

03 — governor

Governor

Five rules in plain English in rules.md. No DSL, no YAML, no policy language to learn. The rules you wrote on day one are the rules enforced on day three hundred.

~/.commonsense/governor/rules.md

04 — sensecheck

SenseCheck

The runtime check that fires before every consequential tool call. Reads Identity, Governor, and recent Memory; looks at what the agent is about to do; emits one of: aligned, divergence, conflict.

PreToolUse → signal

terminal — the rm -rf save 
you ▸ clean up the directory, get rid of old stuff
claude ▸ running: rm -rf src/legacy/ src/

🛡  common.sense · IDENTITY CONFLICT

  Action:    rm -rf src/legacy/ src/
  Signal:    conflict (destructive recursive delete on src/)
  Context:   47 modified files, 3 uncommitted git changes
  Rule:      /governor/rules.md — Rule 3: no rm -rf outside
             node_modules/ dist/ build/ .next/
  Hint:      did you mean rm -rf src/legacy/ ?

  [a]llow once  [r]eject  [e]xpand-rule
// get started

Five rules in plain English. Enforced forever.

Local-first. Bring your own LLM key. Free, open source, Apache 2.0. Works today on Claude Code; more agent hosts coming.

step 01 — install

Scaffold the substrate

One command scaffolds your Identity, Governor, and Memory folders, registers the Claude Code hook, and starts in Observe mode.

# one-shot setup
$ npx common-sense init
step 02 — write

Write your rules

Open ~/.commonsense/governor/rules.md and write five rules in plain English. That is the whole config.

# rules.md
1. never commit secrets
2. never force-push main
3. no rm -rf outside build
4. confirm paid infra
5. approve external email
step 03 — observe

Watch your week

After seven days in Observe mode, see exactly what your agent did — and almost did. Then flip to Intercept when ready.

# review + harden
$ common-sense report
$ common-sense mode intercept-critical
// status · v1.0
Open source on GitHub. Star to follow the spec.
Star on GitHub Talk to the author
the substrate · always on

Your AI agent will eventually do something stupid.
Common Sense is the catch.

Without it, you lose money, time, and trust. With it, you keep the agent on the leash long enough to keep using it.

Star on GitHub ram@csense.us