What is memory-stack?

Memory Stack is a design for giving long-running AI assistants a memory that stays reliable for months, not just a single conversation. The idea behind it is that no one memory technique works well on its own: vector databases drift over time, summaries lose detail, and context windows fill up. Instead of picking one system and hoping it holds, this project splits memory into five separate layers, each with one job, so that if any single layer breaks, the assistant still works with the layers that are left. The five layers are: a raw daily log that records every message exactly as it happened, a short-term rolling summary covering the last few days, a long-term file of durable facts and preferences, a semantic search layer that can look things up across everything stored, and a write-back step that saves new information after each turn. Layers below are treated as the source of truth: if the search layer disagrees with the curated long-term file, the file wins, not the search result. You do not need a vector database to start. The project explains how to get most of the benefit with just three markdown files and a scheduled job that rewrites the short-term summary each night. The full working version, in the memstack folder, is plain Python with no outside dependencies, so it can run anywhere Python 3.10 or later is available. It comes as both a command line tool, with commands to log, search, promote facts, and build context for the assistant, and as a Python library you can import directly. The search layer offers two options: a simple built-in method that compares word overlap, which many projects never need to outgrow, and a slot for plugging in your own vector-based search when you need something more powerful. There are no required accounts, API keys, or paid services baked into the project itself. This project is licensed under MIT, meaning it is free to use, modify, and build on with very few restrictions.