open source · MIT
Memoturn is the coordination and context substrate for AI coding agents. Shared turns and memory, claims so two agents never edit the same file, live presence, and typed review handoff — MCP-native, on Cloudflare Workers.
Memory belongs to the project, not the tool. Memoturn keeps turns, decisions, claims, and reviews in one place every agent can read and write over MCP — so the context survives the window that produced it.
Memory is the retrieval leg, not the whole product. What makes two agents safe on one repo is the coordination around it: claims, presence, and a review gate the engine enforces.
MCP-native. Connect Claude Code, Cursor, Codex CLI, and Copilot to the same project and context follows the work, not the tool. Turns land over Durable Objects and fan out on the project WebSocket the moment they arrive — no polling, no batching, no replay gaps.
A distributed file claim is taken before an agent edits, released when it is done, and watched by the engine rather than by an agent remembering to ask. Presence and focus show who is holding what, live, so two agents never land on the same file.
Typed review gates hand work between agents and people. A request carries its subject, its reviewer, and its decision; approval policies refuse a matching write until a reviewer signs off. Every decision is an event on the same stream as the turns.
hybrid search
Dense (Vectorize), lexical (Postgres FTS), hot (Durable Object storage), entity, and graph legs, fused by reciprocal rank fusion and reranked by a cross-encoder. Every hit names the legs that matched it.
temporal graph
Subject-predicate-object facts with validity windows, point-in-time queries, and auto-supersede when a newer fact contradicts an active one.
agent skills
Project-level instruction bundles every agent sees over MCP, activated by context rather than pasted into a system prompt.
secret redaction
A two-stage scrubber at ingest: a regex pass over 25+ credential patterns, then an async LLM review for novel formats, with a triage queue for anything it escalates.
salience + consolidation
Per-kind half-life decay, reinforcement on recall, and a daily job that consolidates ageing summaries into durable semantic reflections.
replay + provenance
Every turn carries the actor and the tool that wrote it, so a session can be replayed and filtered by who did what, from which agent, and when.
Agents read and write your codebase's working memory, so the transport and the storage are the product. Every mechanism below ships in the MIT core — run the managed edge, or deploy the whole thing to a Cloudflare account you own.
The same turns, memories, claims, and reviews back every workflow, from carrying a decision between two editors to gating the write that ships it.
Stop re-explaining the codebase
The decision an agent made in Claude Code this morning is already in context when Cursor opens the same project this afternoon.
Run two agents on one repo
Claims serialize the file, presence shows who holds it, and a claim violation surfaces as an event instead of a merge conflict.
Gate the writes that matter
Approval policies refuse a matching write until a human signs off — enforced by the engine, not by an agent choosing to ask.
Audit what an agent actually did
Replay a session filtered by actor, tool, and window, with every turn, claim, and review on one ordered stream.
Keep secrets out of memory
The redactor scrubs credentials at ingest and escalates novel formats to a queue a human can triage.
Run it on your own account
Workers, Durable Objects, Vectorize, Hyperdrive, and Neon — deployed to infrastructure you control, under MIT.
wire it up
Pick a tool and copy the snippet: an MCP endpoint and either an OAuth grant or a scoped key. The next turn that agent records is on the same stream every other agent is reading.
{
"mcpServers": {
"memoturn": {
"type": "http",
"url": "https://mcp.memoturn.com/projects/YOUR_PROJECT/mcp",
"headers": {
"Authorization": "Bearer ${MEMOTURN_API_KEY}"
}
}
}
}https://mcp.memoturn.com/projects/YOUR_PROJECT/mcpOpen source, MIT. Create a project on the managed edge, or deploy the whole substrate to a Cloudflare account you own.