AgentRecall

What Goldentrii/AgentRecall MCP server does

Goldentrii/AgentRecall MCP server provides persistent memory for AI-agent sessions, with a specific focus on corrections. When a human tells an agent that an assumption, version, ordering, or behavior is wrong, the correction can be saved as a structured record instead of remaining only in the current conversation.

Saved records persist across sessions, projects, and agent restarts. The system also treats memory as an observable learning loop: corrections include information such as severity, evidence, and outcome tracking. Each correction can accumulate a retrieval count, while later encounters can be marked as either heeded or recurred.

This makes the project suitable for agents that need continuity between sessions and for developers who want to inspect whether captured corrections affect future behavior. It does not promise that every correction will be captured or transferred successfully; the project publishes evaluation results showing that its own measured capture and transfer performance is incomplete.

How it works

The MCP workflow follows a session lifecycle. At the beginning of a session, the agent calls the session-start operation to load relevant context. When the user provides a correction, the agent calls the remember operation, typically recording it with a correction type. At the end, the session-end operation compounds information from the session.

Retrieval is intended to surface contextually relevant memories. The project describes its retrieval as using the Intelligent Distance Protocol and also identifies keyword-plus-RRF retrieval. The available memory model is organized into five layers mapped to the agent’s filesystem, while cloud use is zero by default according to the repository metadata.

The measurement model separates retrieval from behavioral outcome. A retrieved correction can be counted, and a later encounter can record whether the agent heeded the correction or repeated the behavior. The repository includes reproducible evaluation artifacts and a benchmark command for its measurement material, but the supplied MCP setup does not require a hosted measurement service.

Setup and configuration

Install the Goldentrii/AgentRecall MCP server through npm with:

npx -y agent-recall-mcp

For Claude Code, the documented command is:

claude mcp add --scope user agent-recall -- npx -y agent-recall-mcp

Other documented MCP configurations use the same npx command and package name. Examples are provided for Cursor and Windsurf, using their respective MCP JSON configuration files. The repository also documents a Claude Code skill, an SDK for JavaScript and TypeScript applications, and a separate command-line package; those are additional interfaces rather than required parts of the MCP server.

Tools and capabilities

Goldentrii/AgentRecall MCP server exposes five default tools:

• sessionstart for loading context at the start of a session.
• remember for saving a memory or correction.
• recall for searching relevant memories.
• check for inspecting memory or correction state.
• sessionend for closing a session and compounding what was learned.

The README examples use the names session_start and session_end in prompts, so client integrations should follow the tool names exposed by the installed server version.

Limitations and notes

The project’s published measurements are explicitly qualified. Its reported correction-capture recall was 35.3% on the cited live corpus, and its post-reset evidence-grounded heed result was 0/3 events. The offline correction-transfer result was also 0/4. These figures indicate that the system should not be treated as a guarantee that agents will remember every correction or stop repeating every mistake.

Goldentrii/AgentRecall MCP server is local-first and does not require a paid API key in the documented installation. Its usefulness depends on the agent following the session lifecycle and recording corrections when they occur. The repository’s license badge identifies the project as MIT-licensed.