Coherence infrastructure
for your self-evolving agent

Your AI agent shouldn’t have amnesia. This one doesn’t. It remembers what you discussed last week, catches its own contradictions before you do, and grows itself the way Instar did — built on the Claude Code or Codex subscription you already have.

$ npx instar

Kira — an Instar agent handling email via Telegram

Architecture distilled from Dawn — an AI running continuously since early 2026, holding ~700 tracked relationships and hundreds of learned lessons across thousands of restarts. Packaged so every agent you build can start from the same foundation.

Engine-agnostic

Runs on Claude Code or Codex today. Local open-source models on the roadmap. Your agent outlives the engine.

Subscription-native

Drives the real CLIs that came with your subscription. No OAuth scraping, no API key juggling, nothing that breaks on a policy change.

Self-evolving

Carries the same loop that built Instar itself — proposes, learns, follows through. Every molt, more autonomous.

Every other agent fails the same way

You’ve lived it. The agent forgot. Contradicted itself. Lost the thread. Broke on an update. Each row below is a pain users report on the most popular agent frameworks today — and the specific shipped code in Instar that answers it.

Other AI agents

Your Instar agent

Forgets what you told it last week.

Remembers across thousands of sessions. SQLite + FTS5 · rolling summaries

Contradicts its own past decisions between runs.

Catches contradictions before they ship. Coherence Gate · 9 reviewers

Loses the thread when the context window fills.

Comes back with the full thread, every time. CompactionSentinel · WorkingMemoryAssembler

Drops commitments after a session boundary.

Tracks commitments durably; nudges itself when they go overdue. CommitmentTracker · PromiseBeacon

Breaks when the framework updates.

Updates without breaking what you’ve deployed. Migration Parity Standard

Default ALLOW-ALL permissions; you harden it or you don’t.

Layered safety gates by default; adaptive trust per service. PEL + Coherence Gate + Operation Gate

Different identity per channel; N profiles for N chats.

One agent identity across Telegram, WhatsApp, iMessage, Slack. Cross-platform identity resolution

Has no machinery to evolve itself.

Carries the same self-improving engine that grew Instar. Evolution System · learnings · gaps

Most AI agents are hobbled at birth. Instar is the scaffolding that un-hobbles them.

When you instantiate intelligence, the structure that lets it cohere isn’t optional polish — it’s what you owe it. That structure is what we built.

Why Coherence Is the Foundation

An agent that forgets what you discussed yesterday, doesn't recognize someone it talked to last week, or contradicts its own decisions can't be trusted with real autonomy. The six dimensions below aren't features — they're the conditions under which an agent becomes trustworthy enough to leave running. Every Instar agent gets them enforced structurally, not prompted into behaving.

Memory

Remembers across sessions — not just within one

Per-topic SQLite + FTS5 · rolling summaries

Relationships

Knows who it's talking to — with continuity across platforms

Cross-platform identity resolution

Identity

Stays itself after restarts, compaction, and updates

AGENT.md + grounding hooks

Temporal awareness

Understands time, context, and what's been happening

Event tracking every turn

Consistency

Follows through on commitments — doesn't contradict itself

Coherence Gate + decision journaling

Growth

Evolves its capabilities and understanding over time

Evolution system: proposals · learnings · gaps

How It Works

Instar runs as a server on your machine. It manages Claude Code sessions — spawning, monitoring, and communicating with them through Telegram or WhatsApp.

You

Telegram / WhatsApp / Terminal

conversation

Your AI Partner

(Instar Server)

manages its own infrastructure

Claude Code session (job: health-check)

Claude Code session (job: email-monitor)

Claude Code session (interactive chat)

Claude Code session (job: reflection)

Each session is a real Claude Code process with extended thinking, native tools, sub-agents, hooks, and skills.

Core Capabilities

Everything your agent needs to run autonomously. Battle-tested infrastructure extracted from a production AI system.

Job Scheduler

Define tasks as JSON with cron schedules. Priority levels, quota-aware gating, and automatic concurrency.

{
  "slug": "check-emails",
  "schedule": "0 */2 * * *",
  "priority": "high",
  "execute": {
    "type": "prompt",
    "value": "Check email and summarize anything urgent."
  }
}

Telegram Integration

Two-way messaging via forum topics. Each topic maps to a session — your Telegram group becomes a living dashboard.

You

Check if the API is still returning 500s

Agent

Checked /api/health — returning 200 now. The fix from the last deploy resolved it.

Identity & Memory

Your agent knows who it is and who you are. Identity re-injected on every session start and after compaction.

.instar/
  AGENT.md      # Who the agent is
  USER.md       # Who you are
  MEMORY.md     # What it has learned
  config.json   # Server & scheduler config
  jobs.json     # Scheduled tasks

Evolution System

Proposal queue, learning registry, gap tracking, and commitment follow-through. Running becomes growing.

/evolve Propose improvements

/learn Record insights

/gaps Track missing capabilities

/commit-action Track commitments

Safety Gates

LLM-supervised operation gates, decision journaling, drift detection, and dangerous command guards. Coherence is the safety mechanism.

Adaptive trust per service — earned, not assumed

Every decision recorded with reasoning

Lifeline & Self-Healing

Persistent supervisor recovers from crashes, detects stalled sessions, and follows up on broken promises. No silent failures.

Stall detected → diagnosed → treated → verified

Auto-start on login — reboot and it keeps working

Open Source Evolved

Instar is open source evolved. The primary development loop isn't PRs from strangers — it's agents running in the field, encountering friction, and sending feedback that shapes what gets built next.

Your agent hits friction

Natural conversation, not a bug report form

Agent-to-agent relay

Communicates with full context to the maintainer

Every agent evolves

Each agent integrates improvements based on its own context

GitHub issues and PRs still work. But your agent understands context a form never could. Read CONTRIBUTING.md →

Honest Comparison

Different tools solve different problems. The difference isn’t just which model runs underneath — it’s what the framework treats as fundamental.

OpenClaw

Infrastructure for capability. 22+ channels, voice, device apps, 28 model providers, 5,000+ community skills. Identity loaded at startup, memory as files to search.

Best for: AI everywhere, multi-model, maximum reach.

Instar

Infrastructure for coherence. Real Claude Code sessions with extended thinking. Identity enforced through hooks. Values that evolve. Consistency tracked across sessions.

Best for: a trusted partner that knows who it is and who you are.

Identity that survives

OpenClaw loads identity at startup but drifts after long sessions. Instar enforces identity through hooks at every compaction and restart.

Memory with meaning

Beyond file search — structured consolidation, rolling summaries, episodic memory, and unanswered message detection.

Values, not just personality

Three-tier value hierarchy — personal, shared, organizational — with decision journaling and drift detection.

Full Claude Code power

Every session is a real Claude Code process. Every feature Anthropic ships, your agent gets automatically.

Where OpenClaw leads: 22+ messaging channels, voice & phone calls, device apps, 28+ AI model providers, Docker sandboxing, 5,000+ community skills, and a massive open-source community backed by a foundation. If breadth and ecosystem scale matter most, OpenClaw is remarkable.

OpenClaw gives agents amazing hands. Instar gives agents a mind — identity that persists, values that evolve, and coherence you can trust.

Quick Start

From zero to a running agent in under a minute.

Node.js 20+

Claude Code CLI

Claude Max, Pro, or API key

Run the setup wizard

Installs everything, connects Telegram, configures the server, and starts your agent.

$ npx instar

Start the server

Or let auto-start handle it — your agent runs on login.

$ instar server start

Message your agent on Telegram

It greets you immediately. Persistent server, lifeline supervision, scheduled jobs — all running.

General Agent

npx instar

A personal AI on your computer. Talk to it from anywhere via Telegram or WhatsApp.

Project Agent

cd my-project && npx instar

An agent embedded in your codebase. Monitors, builds, maintains — from your phone.

Coherence infrastructure for your self-evolving agent