One markdown source. Every AI tool reads from here. Edit once, the per-tool adapters keep the rest in sync — no more N copies that drift.
CLAUDE.md, GEMINI.md, .aider.conf.yml, .cursorrules, AGENTS.md… seven files, all your context.
Maintaining N copies of "who I am, how I work, what we decided" across every assistant is a daily tax. They drift. They disagree. They're impossible to keep current. aimem fixes this by inverting the relationship: there is one canonical store, and the tools read derivative views of it.
Plain markdown at the bottom. Research-grounded structure on top. A build step in between that handles every tool's quirks.
Edit one file under docs/, config/, notes/… Everything else stays in sync because everything is generated.
One build script emits AGENTS.md, CLAUDE.md, GEMINI.md, Aider config — exactly what each tool reads.
Six unified memory kinds — documentation, semantic rules, workflows, event logs, reflections, references.
Runtime briefing, search, log-turn, write-proposal via Model Context Protocol. Live memory, not just static files.
Compact-mode adapters strip Why and How rationale at build. Sources stay verbose for humans, adapters stay terse for context windows.
Hand-curated, single-author, public so others can fork. Used in production by Lx8 Labs across every project.
aimem isn't an opinion. It's a working implementation of seven papers, in markdown.
Clone, build, symlink. From there: edit any markdown file and the tool-specific adapters regenerate.