Atlas Memory

The token GPS for AI coding — Cursor's where-to-look / what-to-remember layer.

Stop dumping the monorepo into context. Atlas understands the project via indexes, disambiguates the right files, remembers decisions, and saves exploration tokens on the next turn.

mempalace-index  →  MemPalace (optional long-term memory)
graphify-index   →  Graphify or Mind Map (optional; pick one)
project-cache    →  real files (always)

Two commands from zero to a working router:

atlas init                        # creates the indexes and indexes your source tree
atlas connect --editor cursor     # also: windsurf, vscode, zed, codex, claude
# ~/.cursor/rules/atlas.mdc + MCP atlas-mcp — then reload MCP

Prove token savings:

atlas bench --fixture   # synthetic monorepo: 99%
atlas bench --real      # the Atlas repo (needs a source checkout): 78%

Both suites include negative cases, where the correct answer is for the router to stay silent instead of returning plausible-looking noise.

Local daemon (any AI editor):

atlas daemon          # HTTP on 127.0.0.1:8765
atlas-mcp             # stdio MCP

---

Why use Atlas (or migrate to it)

The problem with “just chat with the repo”

What happens today	Cost
Agent greps the whole tree every session	Tokens, time, noise
Last week’s architecture decision lives in an old thread	Rework and contradictory advice
Huge monorepos (or multi-engine trees) overwhelm context	Wrong files, hallucinated paths
“Memory” tools fight each other (two graphs, three “search me first” rules)	Unpredictable agents
Secrets accidentally pasted into notes	Security risk

Your model is not the bottleneck. Orientation is.

Why Atlas

1. One protocol, every project — The agent always asks the same five questions in the same order. No improvisation.
2. Router, not another brain — Atlas doesn’t replace Cursor, Claude, or MemPalace. It tells them where to look.
3. Works at day one, scales later — project-cache alone is useful. Add MemPalace and Graphify when you need them.
4. Safe by default — Typed checkpoints, secret scanning, code-wins-over-stale-memory.
5. Built for real repos — Scoped graphs instead of one mega-index of 30k files.
6. Migratable — atlas migrate / atlas onboard turn README/ADRs and loose Cursor rules into Atlas layout.

Who should migrate

• You already use Cursor (or Claude Code / Codex) on more than one project
• You’ve repeated “we decided X last month” more than twice
• Grep-first workflows feel slow or wrong on a large codebase
• You tried MemPalace / Graphify / Mind Map and want one hierarchy, not three competing rules

Who can wait

• Greenfield toy apps with a handful of files and no lasting decisions — start with atlas init when the project grows.

---

60-second install

pip install atlas-code-memory

cd ~/code/your-app
atlas init --global-rule
atlas onboard
atlas doctor

From source instead:

git clone https://github.com/Ygonet1986/atlas-memory.git
cd atlas-memory
pip install -e .

Open the project in Cursor. The Atlas rule steers the agent automatically.

---

Migration guide

Bring an existing project into Atlas Memory without throwing away README, ADRs, or Cursor rules.

One command

cd ~/code/legacy-app
atlas migrate -C .
atlas doctor -C .

Optional flags:

atlas migrate --dry-run          # plan only (no writes)
atlas migrate --no-import        # skip README/ADR seeding
atlas migrate --global-rule      # also install ~/.cursor/rules/atlas.mdc
atlas migrate --hooks            # install git post-commit stale hook

What migrate does

1. Renames legacy Cursor memory rules → .cursor/rules/atlas.mdc
2. Bootstraps missing indexes / skill / .atlasignore (atlas init)
3. Rewrites Portuguese field labels to English (endereço→path, escopo→scope, …)
4. Ensures an AGENTS.md Atlas section when Cursor rules already exist
5. Seeds project-cache + drawer stubs from README / docs/adr (atlas import)
6. Writes .cursor/atlas-migrate-report.md with next steps

After migrate

1. Read .cursor/atlas-migrate-report.md
2. Fix anything atlas doctor reports
3. Review .cursor/atlas-import/*.drawer.md → atlas checkpoint --write --mine
4. Register scopes: atlas graph add app --scope app (never the monorepo root)
5. Optional adapters: MemPalace and Graphify or Mind Map

English is the canonical language for Atlas indexes and CLI. Parsers still accept older Portuguese labels; new writes use English only. CLI aliases: --escopo → --scope, --descricao → --description.

Full walkthrough (bare repo, Cursor legacy, MemPalace/Graphify, monorepos, teams, rollback): docs/migration.md.

---

How it works

1. mempalace-index   → which memory wing/room?
2. MemPalace         → what did we decide / learn?
3. graphify-index    → which code graph?
4. Graphify|MindMap  → how do symbols connect?
5. project-cache     → which file to open?

Missing layer → skip. Never invent memory hits.

---

Commands (0.3)

Command	Purpose
atlas init	Bootstrap indexes
atlas onboard	Import README/ADRs + onboarding brief
atlas doctor	Diagnose setup
atlas migrate	Migrate existing project → Atlas Memory
atlas route "…"	JSON recall plan for a question
atlas bench	A/B token-proxy proof (grep vs route)
atlas daemon	Local HTTP for any AI editor
atlas connect	Write MCP/rules for Cursor/Claude/generic
atlas graph …	Manage scoped graphs (--scope)
atlas checkpoint --write/--mine	Validate + file drawers by room
atlas life …	Personal memory (wake/remember/sync) + Chat serve
atlas sync export/import	Team bundles
atlas watch	Mark graphs stale on change
atlas eval / --transcript	Index + protocol scores
atlas mcp	MCP stdio server

Life + desktop: docs/life.md · docs/atlas-chat.md

Full docs: docs/index.md · Migration: docs/migration.md

---

MCP (Cursor)

{
  "mcpServers": {
    "atlas-memory": {
      "command": "atlas-mcp"
    }
  }
}

---

Optional adapters

Slot	Options
Memory	MemPalace or none
Code graph	Graphify or Mind Map — never both
File index	Built-in project-cache (required)

---

License

MIT — see LICENSE.