Braingent

Introduction

Why Braingent is Different

Memory you can read, diff, move, and trust — instead of memory hidden inside someone else's database.

Most “memory” features for AI tools share a few traits. They live inside a single product. They store your context as opaque rows in a database you don’t control. They expire on a schedule you didn’t choose. They can’t be diffed, searched offline, or migrated when the tool changes.

Braingent takes the opposite stance.

Memory should be a file, not a feature

Your engineering decisions, reviews, and learnings are some of the most valuable artifacts you produce. They deserve the same treatment as your source code:

  • Plain text that opens in any editor.
  • Versioned in Git, with git log, git blame, and git diff.
  • Diffable so reviewers can spot when memory drifts.
  • Portable — no export step, no proprietary format, no migration.
  • Owned by you, on disk, in a repo you push to a host you trust.

A note hidden inside a chat tool’s session store fails every one of those tests.

Tool-agnostic by construction

Braingent makes one bet: Markdown will outlive every individual AI tool you use this decade. So it stores everything as Markdown with structured YAML frontmatter and lets each agent’s entrypoint file (CLAUDE.md, AGENTS.md, CHATGPT_PROJECT_BRIEF.md, GEMINI.md) point into the same memory repo.

The same record that Claude wrote yesterday is the record Codex will read tomorrow. There is no “Claude memory” and “Codex memory” — there is just the memory.

Retrieval before planning, capture after work

Most chat tools optimize for the conversation. Braingent optimizes for the loop around it:

  1. Before planning — agents search Braingent for prior context, decisions, and known risks.
  2. During work — humans and agents focus on shipping.
  3. After meaningful work — agents capture a durable record and commit it.
  4. Forever after — future sessions inherit instead of starting from zero.

The chat is ephemeral. The records are durable. That’s the contract.

A small CLI on top of files, not a service that owns them

Braingent ships a thin CLI (braingent init, doctor, find, recall, update, qa-generate, task-*) and an MCP server. None of them are required to use Braingent — they speed up things you’d otherwise do by hand.

Delete the CLI tomorrow and your memory still works. The Markdown is the source of truth; the tools are conveniences. That ordering matters.

Privacy is a default, not a setting

Because everything is a file in a repo you control:

  • It runs locally with no network calls by default.
  • It has no hosted account to leak, expire, or be subpoenaed.
  • It puts an explicit list of things it never stores in its starter pack: secrets, tokens, credentials, raw chat transcripts, and sensitive personal data. Your reviewers can grep for violations.

You decide what gets pushed to which remote. You decide who has read access. You decide when to rotate or redact.

Built for multiple agents at the same time

A single Markdown record can be read by Claude in one terminal, Codex in another, and ChatGPT on the web — at the same time, with no sync layer. Multi-agent task files (tasks/active/BGT-NNNN.md) coordinate handoffs between them with nothing more than diffs.

This was built for the way modern engineers actually work: across multiple tools, sometimes in parallel, often switching mid-task.

What this means for you

You don’t have to bet on a single AI vendor’s memory roadmap. You don’t have to lose context when you switch tools. You don’t have to pay a SaaS bill for the privilege of remembering your own decisions.

You write Markdown. Git tracks it. Your agents read it. That’s the whole trick.