Skip to main content

Why Mindr

The problem: agents start every session blind​

Agent amnesia

Modern AI agents work in isolated context windows. When a new session starts, the agent has lost every prior refactor, every "why" decision, and every preference you established. The result: regressive bugs, circular reasoning, and hours of manual re-prompting.

A coding agent's context window resets between sessions. Every time you open a new conversation, the agent re-derives the same things you told it yesterday:

  • What language, framework, and database this project uses
  • How you name functions, files, and tests
  • Why you switched from REST to tRPC three weeks ago
  • Which files have known FIXMEs it should be careful around
  • How to actually build, test, and lint the project

So you re-explain. Or the agent guesses — inconsistently — and you correct it. Again. That re-orientation tax is paid on every session, by every teammate, forever.

The usual workarounds don't hold​

ApproachWhy it falls short
Hand-written AGENTS.md / CLAUDE.mdAccurate on day one, stale by month two. Nobody updates it after every decision.
Longer promptsBurns tokens re-describing the codebase, and still can't recall why past decisions were made.
RAG over the whole repoRetrieves code, not the reasoning, conventions, or debt behind it — the things that aren't in any single file.

How Mindr is different​

Mindr treats memory as a byproduct of the work you already do. A post-commit hook reads each commit and extracts durable, structured memory — then serves it to any agent over MCP or writes it into AGENTS.md / CLAUDE.md.

  • Automatic. No prompting and no manual note-taking. Commit as usual; memory accrues.
  • Structured, not retrieved. Conventions (with confidence scores), decisions (with the triggering commit), debt (aged and severity-ranked), and bug patterns — not raw file chunks.
  • Self-maintaining. Debt is marked resolved when you delete the marker; conventions re-derive as the codebase evolves. The context never goes stale.
  • Agent-agnostic. One MCP server for Claude Code, Cursor, Codex, Aider, Windsurf, Continue, and OpenCode. Bring whatever tool your team uses.
  • Local-first. SQLite by default — zero config, nothing leaves your machine. Move to Remembr later for cross-machine, cross-team persistence.

Before and after​

Before — a fresh session, every time:

You:   We use tRPC for internal APIs, REST for external. camelCase functions,
       PascalCase classes. Tests live next to source as *.test.ts. Auth is JWT
       with Redis refresh tokens. Don't touch billing/invoice.ts:47 — the retry
       logic is broken. Build with pnpm, test with pnpm test...
Agent: Got it. (…until the next session, when you say it all again.)

After — the agent reads it in one call:

=== MINDR CONTEXT ===

[RECENT DECISIONS]
  2026-05-01 — migrate auth to JWT + Redis [keyword]
  2026-04-15 — switch internal APIs to tRPC [keyword]

[COMMANDS]
  Install: pnpm install
  Test:    pnpm test

[CONVENTIONS]
  typescript: functionNames=camelCase(97%), classNames=PascalCase(100%)

[WARNINGS]
  FIXME `src/billing/invoice.ts:47` — retry logic (high, 43d)

=== END CONTEXT ===

When Mindr is a good fit​

  • You use one or more AI coding agents and re-explain your codebase to them often.
  • You want a living AGENTS.md that stays correct without manual upkeep.
  • You care about why past decisions were made, not just what the code says today.

Ready? Head to the Quickstart.