What is aiDimag?

aiDimag is a memory system for AI coding agents. Its command-line tool is called dim.

The problem it solves

AI coding assistants are powerful but forgetful. Every new session starts from zero:

• They re-discover the same architecture you explained yesterday.
• They repeat approaches your team already tried and abandoned.
• They confidently follow a "convention" that was true six months ago but isn't anymore.
• They have no idea which rules are sacred ("never call the production API from a test").

Teams try to fix this with a big CLAUDE.md or a wiki, but those rot. Nobody updates them, and worse — nothing checks whether what they say is still true. A stale instruction is more dangerous than no instruction, because the agent trusts it.

The idea: verified memory

aiDimag treats every piece of knowledge as a falsifiable claim with evidence, not a free-text note. For example:

Claim: "All database access goes through src/db/store.ts; nothing else imports better-sqlite3."

Evidence: the shell command grep -rL better-sqlite3 src --include=*.ts (passes only if the claim holds).

Because the claim is checkable, aiDimag can re-run the evidence as your code changes:

• If the check still passes, the memory stays trusted.
• If the code changes and the check fails, the memory is marked stale — and the agent is told not to rely on it until it's re-confirmed.

That's the whole wedge: memory that proves itself, instead of memory you hope is right.

What you can remember

aiDimag organizes knowledge into kinds:

Kind	What it captures
Decision	A choice you made and why (what you rejected)
Convention	A rule the repo consistently follows
Gotcha	Surprising behavior that cost someone time
Failed approach	Something tried that did not work — so nobody retries it
Architecture	How the pieces fit together
Invariant	Something that must always (or never) hold
Guardrail	A behavioral rule for agents: never, ask first, or always
Skill	A reusable step-by-step procedure (e.g. "how we deploy")
Todo context	Unfinished work plus the context to resume it

Who it's for

• Developers who pair with AI agents and want them to stop forgetting.
• Teams who want a shared, self-correcting brain for a codebase.
• Any AI tool — aiDimag speaks the Model Context Protocol for Claude Code, Cursor, and Copilot, and generates plain context files (CLAUDE.md, .cursorrules, copilot-instructions.md) for tools that don't.

What makes it different

• Local-first. Everything lives in a single SQLite file in .aidimag/ next to your code. No account required to start.
• Human-gated. Nothing becomes active memory automatically — proposals wait in a review queue until you approve them.
• Self-correcting. Memories carry confidence that decays over time and collapses when evidence fails, so trust expires instead of lingering.
• Team-optional. Add a self-hosted sync server when you want a shared brain. No SaaS lock-in.

Next: Core concepts explains the moving parts, or jump to Getting started.