Stele — shared memory for AI coding agents

How it worksIn practicePricingDocsSign inJoin the waitlist

Every agent, fluent in your project.<br>Stele gives the tools you already use a shared memory of your decisions, tasks, and lessons — one record every agent reads before it acts. Switch tools or sessions; the context stays.<br>Join the waitlist Have an invite code? Sign in<br>Invite-only beta · we'll email you the moment a spot opens — no spam.<br>Already invited? Install the CLI$curl -fsSL https://app.stele-ai.dev/install.sh | bashCopy

stele · knowledge graph live<br>Overview<br>Knowledge<br>Tasks<br>Documents<br>Graph<br>Usage<br>componentdecisionrisktask

open7<br>TASK-204P1<br>Rotate the signing key on a schedule

TASK-219P2<br>Backfill missing component anchors<br>1/4

in progress2<br>TASK-198P0<br>Make account deletion atomic + resumable<br>3/5op

TASK-211<br>Guard the complete-summary gate<br>2/3cc

done41<br>TASK-191<br>Atomic claim RPC for tasks

TASK-187<br>Search RPC over the node graph

2hKNOW-118Auth lock guards a race that reached prod once<br>riskreview

1dKNOW-90Hosted assistant streams via the AI Gateway<br>architecture

2dKNOW-76Design system: oklch tokens, one type ramp<br>decisionverified

4dKNOW-52Atomic claim keeps two agents off one task<br>lesson

6dKNOW-31Expand/contract every schema migration<br>goal

Explore the live app

Plugs intoClaude CodeCursorCodexAntigravityCopilotOpenCodeGrok BuildKimi Code CLI

What Stele is<br>A project memory every agent reads — and writes.<br>Stele is a shared ledger: a live record of what your agents know, what they've decided, and what they're working on. Every agent reads it before it answers. Everything useful it learns, it writes back.

One living memory.<br>Decisions, lessons, and risks in a shared record every agent reads before it acts — and writes back as it learns. Not a passive log: it surfaces the right context at the moment a change would repeat a past mistake.

Work the whole team can see.<br>Tasks belong to the project, not a person. Agents claim them atomically, so two sessions never start the same work — and an operator can pick up the open queue and run it end to end.

Continuity across every tool.<br>Plan with one agent, build with another, switch when a quota fills. Claude Code, Cursor, Codex, and more read the same graph — so your project's memory belongs to your team, not to any one tool.

How it works · about a minute<br>Three commands, then it's automatic.<br>No dashboard to configure and no workflow to learn. You set it up once, then keep prompting your agent exactly the way you do today.

01<br>Get your invite & sign in<br>One line installs the stele CLI. Sign in with the email your invite was sent to — no separate account to create.

02<br>Run /stele:start<br>Inside your agent, run it once. Stele wires into your project and builds the first record.

Keep working<br>That's it. Your agent now reads the record before it answers and writes back what it learns.

Copy & run — that's step one<br>terminalCopy<br># installs the stele CLI & signs you in<br>$curl -fsSL https://app.stele-ai.dev/install.sh | bash<br>$stele account login<br>→ signed in ✓

macOS, Linux & Windows · one binary. Sign-in works once you're invited.

In practice<br>The record catches what a fresh agent can't.<br>A small moment that happens in any project with more than one contributor — human or agent. Here's how it goes when the reasoning lives in the record instead of someone's memory.

Catch a returning bugA harmless-looking change reintroduces a problem you already paid for.<br>The linked risk surfaces while the work is happening — not after it ships. The agent pauses and asks instead of guessing.<br>See more examples in the docs<br>your agentstele · on<br>You"This lock looks unnecessary — I'm removing it."

SteleThat lock came from TASK-91 during the auth rework. Removing it reintroduces ⚠ KNOW-118, a race condition that reached production once. Want me to keep the guard and simplify around it instead?

Works with your agents<br>Bring the agents you already use.<br>Stele plugs into the coding agents you work in today — and any MCP-compatible client through the hosted server. Plan with one, build with another, switch when a quota fills; they all read and write the same project graph.

Claude CodeCursorCodexAntigravityCopilotOpenCodeGrok BuildKimi Code CLIany MCP client

More than memory<br>Everything orbits the record.<br>The memory is the center; these are the tools you act on it with — each one sharper because it shares the same graph.

Kept honestA graph that maintains itself<br>Stale entries are flagged, resolved risks retire, duplicates surface, and reversed decisions stop being served as truth — and your agents re-verify facts against the live code, so the record stays trustworthy instead of rotting as it grows.

ProvenanceTrace why anything exists<br>Walk from a decision back to the task, the document, the risk, and the person who made the call. Your agents act on the reasons behind the work, not just the instruction in front of them.

CoordinationTasks the team claims<br>Project-level tasks, claimed atomically so no...