real-time
Every turn, captured live
Durable Objects fan turns out over WebSockets the moment they land. No polling, no batching, no replay gaps.
Durable ObjectsWebSocket
open source · MIT
Real-time AI memory your coding agents share.
Claude Code, Cursor, Codex, and Copilot read and write one project memory over MCP. Hybrid search and a temporal knowledge graph — on Cloudflare's edge.
- License
- MIT licensed
- Protocol
- MCP-native
- Runtime
- Cloudflare Workers
- Tools
- 42 MCP tools
works with
- Claude Code
- Cursor
- Codex CLI
- VS Code Copilot
Your agents shouldn't be amnesiacs. Memory belongs to the project, not the tool.
Today each coding agent keeps its own private context, so the work resets every session and never crosses tools. Memoturn inverts that: one shared substrate per project, written and read over MCP — the same memory whether the next turn comes from Claude Code, Cursor, Codex, or the CLI.
A complete memory substrate, every leg named.
No black boxes. Every leg is instrumented, every score is inspectable. The dashboard surfaces which leg matched and why — the search isn't magic, it's accountable.
cross-tool
One memory, every agent
MCP-native. Connect Claude Code, Cursor, Codex CLI, and Copilot to the same project — context follows the work, not the tool.
MCPstdio + http
hybrid search
5 legs, fused with RRF
Dense (Vectorize) + lexical (Postgres FTS) + hot (DO storage) + entity + graph, fused by reciprocal rank fusion. See which leg matched, every time.
VectorizePostgres FTSRRF
also in the substrate
temporal graph
Subject-predicate-object facts with validity windows, point-in-time queries, and auto-supersede on conflict.
factsedgessupersede
agent skills
Project-level instruction bundles every agent sees over MCP. Contextual activation, no system-prompt bloat.
skillsregistry
coordination
Distributed file claims, peer reviews, live presence and focus maps. Two agents on the same file don't collide.
claimreviewpresence
secret redaction
Two-stage scrubber: a regex pass for 25+ credential patterns, then async LLM review for novel formats.
regexLLM reviewaudit
salience + consolidation
Per-kind half-life decay, reinforce-on-recall, daily consolidation of old summaries into durable reflections.
salienceconsolidationsynthesis
Stop re-explaining your codebase to every agent.
Open source, MIT licensed. Sign up to the hosted service, or clone the repo and deploy your own.