darula-hpp/memgrep

memgrep is a local memory and scheduling system built for people who code with Cursor. It runs entirely on your own machine and gives coding agents a way to remember past work instead of starting fresh every session. The core idea is that agent chats burn time and tokens solving problems that were already solved in an earlier conversation. memgrep fixes this by ingesting past chat transcripts from Cursor, Claude Code, and Kiro, storing them locally in a small database, and letting you search them later with a mix of keyword and meaning based search. You can also manually save notes or playbooks with a simple remember command, then pull them back up mid task using recall. Beyond memory, memgrep adds a coding loop feature. You give it a task, it works through the code with Cursor, checks defined exit conditions, and can automatically open a pull request once the work passes. There is also a job scheduler built on cron, so you can set playbooks to run on a timer, for example a health check every five minutes. A Telegram bot is included so you can trigger a real local Cursor agent from your phone rather than only from a laptop. Everything connects through one MCP server, which is the same protocol used by tools like Claude Code, so any MCP capable agent can plug into memgrep's memory, jobs, and loop features. Optional extras exist for tools such as Jira, Neon, and PostHog, but these only appear once configured with credentials. Setup requires Node.js 18 or newer, and the small local embedding model downloads once on first use so that search works offline afterward. Using the coding loop, Telegram bot, and scheduled jobs does require network access and a Cursor API key. The project is aimed at developers already comfortable with Cursor who want their agent workflows to persist and repeat rather than restart from scratch each time.