uxkero/lastmemory

lastmemory is a memory skill for Claude Code and similar AI coding assistants that gives them a persistent memory of your project without eating up huge amounts of context every time. It keeps a small folder of linked markdown files inside your project, where each file acts like a neuron representing a coding session, a part of the repo, a decision that was made, a failure worth remembering, or a pattern noticed across sessions, connected to each other by links the agent can walk instead of reading everything at once. The README describes four core ideas behind the design: memory is organized by which part of the repo it relates to rather than by a simple timestamped log, files link to each other like a small network, information loads in cheap layers so the full context is rarely dumped in at once, and old or unimportant memories actively get merged, compressed, or dropped over time through a process called a dream pass, rather than growing forever. You use it through slash commands inside Claude Code. Turning it on for a project creates the memory folder, saving a session writes a summary and updates the relevant zones, and asking it a question queries the network and answers with the source cited. A dream command consolidates old memories, and a view command generates an offline HTML graph you can open in a browser to see the whole memory network, colored and sized by how connected and recently used each piece is. The README backs its efficiency claim with a measured benchmark script, reporting that catching up on a sample project costs about 1,435 tokens through lastmemory versus 10,713 tokens to dump every memory file directly, roughly an 86 percent saving, measured with the tiktoken tokenizer on a bundled example project. It draws its underlying techniques from several existing memory research projects, credited openly in the README. The project is released under the MIT license.