AI Optimizer

Reduce token waste and strengthen context focus for AI coding agents with a single setup script.

Quick start

Run from a project root (repo with .git; the script can create docs/ when you choose QMD):

cd /path/to/your-repo
./setup.sh

Non-interactive examples (Cursor CLI):

# Local memory (ICM) + QMD
AGENT=cursor MEMORY_TOOL=icm DOCS_TOOL=qmd bash setup.sh

# Cloud memory (Mem0) + Graphify
AGENT=cursor MEMORY_TOOL=mem0 DOCS_TOOL=graphify MEM0_API_KEY=m0-your-key bash setup.sh

What gets installed

Tool	Role	Setup
RTK	Compress Shell/CLI output before it reaches the agent	Always installed — rtk init -g --agent cursor → ~/.cursor/hooks.json
ICM or Mem0	Persistent cross-session memory (pick one)	ICM: local SQLite, no account (rtk-ai/icm). Mem0: cloud MCP, requires API key (mem0ai/mem0)
QMD or Graphify	Project context (pick one)	QMD: semantic search over docs/**/*.md. Graphify: knowledge graph + Cursor MCP (graphify.serve) for larger codebases/monorepos
tgrep (optional)	Fast indexed regex search on large trees	Prompted during setup with a file-count hint; install via Homebrew or GitHub Releases. Complements Graphify/QMD — not a replacement

Choosing ICM vs Mem0

During setup you pick one memory tool:

• ICM — simple local memory stored in SQLite on your machine. No account or API key. Best default for most users. Configures MCP + ~/.cursor/rules/icm.mdc.
• Mem0 — cloud-hosted memory with semantic search. Requires a free account and API key from app.mem0.ai. Configures MCP + ~/.cursor/rules/mem0.mdc.
• None — skip memory if you only want RTK (+ optional docs tool).

Set non-interactively with MEMORY_TOOL=icm, MEMORY_TOOL=mem0, or MEMORY_TOOL=none.

Back-compat: ENABLE_ICM=yes|no or ENABLE_MEM0=yes|no map to MEMORY_TOOL when MEMORY_TOOL is unset.

Choosing QMD vs Graphify

During setup you pick one documentation/codebase tool:

• QMD — best when your project knowledge lives in markdown under docs/. Registers a collection like <repo>-docs, runs qmd embed, and installs a post-commit hook to keep embeddings fresh.
• Graphify — best for larger codebases or monorepos where you need relationship queries across code. Installs graphifyy[mcp], registers Cursor MCP (python -m graphify.serve …/graphify-out/graph.json), writes ~/.cursor/rules/graphify.mdc, and installs a post-commit hook to rebuild the graph after code commits. Build once with graphify . (or BUILD_GRAPHIFY=yes).
• None — skip both if you only want RTK (+ optional memory).

Set non-interactively with DOCS_TOOL=qmd, DOCS_TOOL=graphify, or DOCS_TOOL=none.

Optional tgrep (large-repo code search)

After RTK/memory/docs setup, the script estimates how many files are in the tree and asks whether to install microsoft/tgrep — trigram-indexed grep that shines on monorepos (often 10k+ files). Small repos get a hint that built-in ripgrep is usually enough; you can still opt in.

• Install: INSTALL_TGREP=yes (or answer Yes at the prompt)
• Index during setup: BUILD_TGREP_INDEX=yes
• Background server: TGREP_START_SERVE=yes (otherwise run tgrep serve . yourself)

Adds .tgrep/ to .gitignore, writes ~/.cursor/rules/tgrep.mdc, and documents usage in AGENTS.md when enabled.

How setup.sh works

1. RTK — installs and runs rtk init for your agent so command output is compressed before the model sees it.
2. ICM or Mem0 (optional, mutually exclusive) — configures persistent memory via MCP and writes agent rules.
3. QMD or Graphify (optional, mutually exclusive) — indexes project docs or builds a knowledge graph so the agent can retrieve context without reading entire files.
4. AGENTS.md / GEMINI.md — appends an "Optimization Utilities" section documenting what was configured.

Environment variables

Variable	Values	Description
AGENT	cursor, github-copilot, antigravity	Target AI agent
MEMORY_TOOL	icm, mem0, none	Memory backend (local vs cloud)
MEM0_API_KEY	m0-...	Mem0 Platform API key (only when MEMORY_TOOL=mem0)
DOCS_TOOL	qmd, graphify, none	Documentation/codebase context tool
BUILD_GRAPHIFY	yes / no	Build graphify-out/graph.json during setup (when DOCS_TOOL=graphify)
INSTALL_TGREP	yes / no	Install optional tgrep (skip prompt)
BUILD_TGREP_INDEX	yes / no	Run tgrep index . when tgrep is installed
TGREP_START_SERVE	yes / no	Start tgrep serve . in the background after setup
AUTO_INSTALL_PREREQS	yes / no	Auto-approve system package installs
SKIP_AGENT_CHECK	yes / no	Skip agent install verification (e.g. Docker)

How the tools affect agent work

RTK (token-efficient inputs)

rtk init -g --agent cursor configures Cursor-side hooks (~/.cursor/hooks.json). Large CLI/Shell outputs are compressed before the model sees them.

ICM (local cross-session memory)

When MEMORY_TOOL=icm, the script installs ICM, runs icm init --mode mcp + icm init --mode skill, then explicitly merges icm into ~/.cursor/mcp.json (<abs-path>/icm serve) so it coexists with Graphify/other MCP servers, and ensures ~/.cursor/rules/icm.mdc. The agent can icm_memory_recall / icm_memory_store (or icm recall / icm store) — no cloud account required.

Mem0 (cloud cross-session memory)

When MEMORY_TOOL=mem0, the script configures the Mem0 MCP server in ~/.cursor/mcp.json and adds ~/.cursor/rules/mem0.mdc. Requires MEM0_API_KEY. The agent can search_memories / add_memory across chats and devices.

QMD (markdown docs)

When DOCS_TOOL=qmd, the script registers docs/**/*.md in a QMD collection, embeds content, and hooks post-commit for re-indexing. Prefer qmd search / qmd query over pasting whole files.

Graphify (codebase relationships + MCP)

When DOCS_TOOL=graphify, the script installs graphifyy[mcp], runs graphify hook install (post-commit graph rebuild), graphify cursor install, merges a graphify entry into ~/.cursor/mcp.json (using a Python interpreter with the [mcp] extra), and writes Cursor rules. Run graphify . once (or set BUILD_GRAPHIFY=yes), restart Cursor, then prefer MCP tools (query_graph, get_neighbors, shortest_path) or CLI graphify query / graphify path.