Jevmem: Automated Project Memory for AI Coders

jevmem is a tool that automatically records project decisions and constraints from AI coding chats into a persistent JEVMEM.md file. It utilizes the Jev API for fast, low-cost decision-making and includes features like secret scrubbing and versioning for superseded memories. While primarily optimized for Claude Code, it also supports Cursor and Codex through various integration methods.

Key Points

• Automatically captures and versions project decisions, constraints, and bugs into a local Markdown file to maintain long-term context.
• Integrates seamlessly with Claude Code via hooks and supports Cursor and Codex through the Model Context Protocol (MCP).
• Prioritizes speed and cost-efficiency, outperforming major LLMs like GPT-6 and Claude Opus in decision latency while maintaining competitive accuracy.
• Includes built-in privacy features that scrub sensitive information such as API keys, tokens, and emails before data leaves the local machine.
• Provides a CLI for auditing, searching, and manually refining memories, including the ability to re-train decision thresholds based on user feedback.

Sentiment

The overall sentiment is mixed but skeptical. Commenters generally agree that persistent memory for coding agents is a meaningful problem, and the author's clarifications make the project easier to understand, but the dominant reaction questions the README, the reliability of automatic relevance classification, and the privacy implications of routing development context through Jev. The thread is more constructively critical than hostile: the strongest objections focus on adoption risks and product clarity rather than dismissing the underlying need outright.

In Agreement

• Coding agents do lose important project decisions across sessions, and a persistent repo-local memory file addresses a recognizable workflow problem.
• Keeping decisions, constraints, bugs, and open questions in a readable Markdown file can be preferable to opaque, harness-specific memory stores.
• Superseded decisions can be useful as history if they are preserved for traceability while excluded from active prompt injection.
• A save/skip gate with typed memory categories is a more plausible approach than logging every turn indiscriminately.
• Several commenters are curious about comparing Jevmem with adjacent project-memory and decision-tracking tools, suggesting the problem space has real demand.

Opposed

• The README fails to clearly explain what the tool is for or why it is better than native memory features, making the project feel overdocumented but underexplained.
• Automatic memory systems may not reliably distinguish durable project knowledge from trivial implementation details, causing context pollution over time.
• Sending coding conversation context and memory lines to a young hosted vendor creates privacy and trust concerns, even with some redaction in place.
• Some commenters question whether Jev offers enough advantage over a cheap constrained model or local classifier with a schema-bound tool call.
• The broader AI-tooling framing strikes some readers as hype-driven, with skepticism toward another layer of agent infrastructure.