Stele
Shared, persistent project memory for AI coding agents, reachable over MCP by Claude Code, Cursor, Codex and other clients.
At a Glance
Enough to run a real project on. No card needed.
Engagement
Available On
Listed Oct 2026
About Stele
Stele is a hosted project memory built by Serkan Yersen that lets AI coding agents share one record of what a project has decided, tried, and is working on. Agents read the record at the start of a session and write back as they work. It is currently in open beta and connects to Claude Code, Cursor, Codex, and any MCP-compatible client.
What It Is
Stele is a hosted graph of knowledge (decisions, lessons, risks, goals), tasks, and components, linked together and addressable by stable IDs such as TASK-12 or KNOW-4. Agents reach it over MCP or through a local plugin and CLI. Instead of each session starting cold, an agent pulls relevant history into context and files what it learns afterward.
How the Workflow Runs
Setup is an installed CLI (macOS, Linux, and Windows through WSL), signing in to create an account, and running /stele:start inside the agent to build the first record. After that, context is recalled automatically in each session. Agents claim tasks atomically so two sessions do not start the same work, and leave handoff notes when a session ends mid-flight. An Operator reference loop pulls open tasks, plans them, executes inline or in isolated worktrees, verifies results, and reports back.
Keeping the Record True
The site describes a graph that flags drifted entries for review, retires reversed decisions and resolved risks, surfaces duplicates for merging, and lets agents re-verify facts against live code. When a proposed change would repeat a past mistake, the record surfaces the earlier attempt before work begins. An assistant answers questions about project state from the live graph, and a backfill step can seed the record from an existing repo.
Data and Access
The vendor states that project graphs are private to members unless published, source code stays on the user's machine, data is not sold, and it is not used for model training without opt-in. Users can export or delete their data. Public projects can be browsed without an account, and Stele can also be added to ChatGPT from its app directory. An open, SQLite-backed self-hosted version is described as planned.
Community Discussions
Be the first to start a conversation about Stele
Share your experience with Stele, ask questions, or help others learn from your insights.
Pricing
Free
Enough to run a real project on. No card needed.
- Unlimited public projects
- 1 private project
- 500 knowledge writes / mo
- 5,000 stored nodes
- Search, recall & hosted MCP
Indie
Side projects you'd rather keep private
- Everything in Free, plus
- 3 private projects
- 1,500 knowledge writes / mo
- 15,000 stored nodes
- 500 assistant turns / mo
- Email support
Pro
Solo devs whose project became the main thing
- Everything in Indie, plus
- Unlimited private projects
- 3,000 knowledge writes / mo
- 50,000 stored nodes
- 5 members per private project
- 1,500 assistant turns / mo
- Monthly knowledge review (coming soon)
- Custom domain
Team
When you bring people on to build with you
- Everything in Pro, plus
- Unlimited members & projects
- 5,000 assistant turns / mo, pooled across the team
- Biweekly knowledge review (coming soon)
- Roles + audit log
- Cross-project search
- Dedicated support
Enterprise
Large repos, custom limits, an SLA
- Custom limits
- SLA
- SSO and self-hosting available as add-ons
Capabilities
Key Features
- Shared persistent memory graph for coding agents
- Knowledge nodes for decisions, lessons, risks, and goals
- Task queue with atomic claiming and handoff notes
- Automatic context recall at session start
- Reversed decisions retired and drift flagged for review
- Operator loop that plans, executes, and verifies tasks
- Provenance tracing from decisions to tasks and documents
- Assistant for querying project state
- Backfill from existing repos and docs
- Public project exploration
- Data export and deletion
