Adaptive Recall: Agent's persistent memory becomes an MCP module

A quiet Show HN that points to the real question: where does the memory of your agents live, and who controls it?

The fact

On July 12, 2026, Adaptive Recall was presented as Show HN: a persistent memory layer for AI agents, exposed via MCP. Goal: to provide clients (Claude, Cursor, IDE agents…) with externalized long-term memory, with semantic search and scope management.

Our take

Agent memory is no longer an R&D topic - it's a product layer. By externalizing via MCP, Adaptive Recall bets that memory is an object independent of the client: it survives model changes, session resets, and tool switches. It's the sketch of a new primitive.

Under the hood

The pattern is classic: embeddings for indexing, scopes by "project" or "user", hybrid retrieval (vector + keyword), recall injected into the context via an MCP tool. What's new is where the memory lives: outside the client, portable, queryable by any agent compatible with MCP. Immediate corollary: it's also an attack surface - see "State of MCP Security 2026" published the same day.

So what

For an agent builder, two questions: (1) where should the memory of your workflow live - in the harness, in the model, or in a third-party MCP service? (2) who inherits control if you externalize? The answers pull the architecture in opposite directions.

To watch

Anthropic or OpenAI announcing a standard "MCP Memory" profile; cyber actors starting to audit persistent memory servers; native integration into Cursor or VS Code Agent.