Jevmem

What the jevmem MCP server does

The jevmem MCP server maintains a project memory file named JEVMEM.md in the repository. It records selected decisions, rules, bugs, to-dos, and approaches that were tried and did not work. When a later prompt relates to saved entries, jevmem supplies the relevant lines to the coding agent.

The file is plain text and can be committed with the project, so a team can share the same recorded context through version control. When a decision changes, the earlier entry is marked as superseded rather than removed from the file. Entries also contain metadata such as an identifier, timestamp, and confidence value in comments.

How it works

During a conversation, jevmem sends a yes-or-no retention question with probabilities to Jev, a model from TypeSafe AI. If the content is worth keeping, jevmem writes a line to JEVMEM.md. On a later session, it selects lines relevant to the current prompt and passes them back to the agent.

The guard checks commands before execution and file edits before they occur. If a proposed action may conflict with a saved rule, Claude Code can ask for confirmation. The guard is described as a backstop based on shared words between the rule and the proposed call, not as a sandbox or complete security boundary.

Setup and configuration

Installation requires a TypeSafe API key, which is used for jevmem's requests, and Claude Code 2.1.273 or newer for the Claude Code plugin workflow. Install the CLI globally with npm, save the key using jevmem key, and enable the tool from the project directory with jevmem enable. The jevmem doctor command checks the setup.

The README also documents npm-based initialization for Claude, Cursor, Codex, Claude Desktop, or all supported targets. Existing CLAUDE.md content can be reviewed with jevmem import; adding --apply writes the proposed additions.

Tools and capabilities

• Automatically save selected project context to JEVMEM.md.
• Recall memory relevant to a later prompt.
• Preserve superseded decisions for history while excluding them from current context.
• Check commands and file edits against saved rules.
• Provide MCP access for Cursor, Codex, and Claude Desktop.
• Run diagnostics with jevmem doctor.

Limitations and notes

The README's reported measurements come from small, non-independent test runs, so they should not be treated as general benchmarks. A manually maintained CLAUDE.md performed similarly in the described tests and did better for a convention that was not referenced by the prompt. Rules that should apply to every task still belong in CLAUDE.md.

The guard is not a sandbox. It can produce questions for calls that do not break a rule, and its documented matching approach depends on words shared by a saved rule and a tool call. Retrieval time and cost increase with the size of JEVMEM.md.

The jevmem MCP server depends on the external TypeSafe API, and the README states that requests are billed to the configured TypeSafe account. Automatic saving is described for Claude Code and for Codex while jevmem watch is running; Cursor, Codex, and Claude Desktop access it over MCP.