MCP is becoming the universal interface for agent memory

For a while, "agent memory" meant something different depending on which framework you were standing in. A LangGraph app had a checkpointer and a store. A Claude Code session had whatever context fit in the window. A Cursor project had nothing beyond the current chat. Every framework was solving the same problem — how does an agent carry what it learned into the next task — with its own incompatible answer.

That's starting to change, and the reason is the same reason editors stopped needing a bespoke autocomplete engine per language: a shared protocol underneath. The Language Server Protocol let one language server work in any editor. The Model Context Protocol (MCP) is doing the same thing for agent tools — including memory. An MCP server exposes tools once; any MCP-compatible client can call them, regardless of what framework built the client.

The signal is already loud

Three things happened in the same few weeks that make this obvious in hindsight. LangChain — one of the largest agent frameworks — now documents that its Fleet product can be "extended with any MCP server," and ships an official langchain-mcp-adapters package so a LangGraph agent can call an MCP tool as naturally as a native one. LangSmith is running paid search campaigns for the keyword "agent memory" outright — proof that a well-funded platform considers this a category worth buying, not a feature worth bundling quietly. And independent tools like ctx are building agent-history indexing as standalone CLIs rather than framework plugins, because framework lock-in is exactly the constraint developers want out of.

Where threadctx fits

threadctx picked the MCP-first design from day one, for a specific reason: memory is only valuable if it outlives the tool that created it. An agent that learns something in Claude Code on Monday and a different agent that picks up the same repo in Cursor on Tuesday should not be starting from zero — and increasingly, neither should a LangGraph agent that touches the same codebase on Wednesday. Two tools — memory_write and memory_query — over one protocol, usable by whatever agent your team happens to be running.

See it wired into a LangGraph agent, or into Claude Code and Cursor — same server, same two tools, every time.