Skip to content

Home

Jump to bottom
Brian Vasquez edited this page May 20, 2026 · 36 revisions

A.I.M. (Actual Intelligent Memory)

"Treat your AI like a bot, not an oracle. Built by a gamer, for the trenches."

☕ Support the project: Buy Me a Coffee

License: MIT Version Platform

A.I.M. is an open-source engineering OS designed to cure the "Amnesia Problem" of autonomous AI agents.

🛑 The Problem: The 50-Turn Spaghetti Code

If you let an AI agent code autonomously, it is brilliant for the first 10 turns. By turn 50, its context window is overflowing, it forgets your architectural rules, it hallucinates answers, and your repository degrades into unmaintainable spaghetti code.

The industry's answer is to build a bigger context window. That's giving a distractible worker a bigger desk.

A.I.M.'s answer is a strict Standard Operating Procedure, a memory sieve, and a leash.

🚀 The Solution: A Sovereign Exoskeleton

A.I.M. wraps around your CLI agent (Gemini CLI / OpenCode) and forces it to act like a disciplined Principal Engineer. It externalizes the agent's memory into a local LanceDB vector pool, automates its Git lifecycle through atomic deployments, and prunes its thoughts while you sleep.

🔥 Core Capabilities

A.I.M. provides a massive suite of tools to control, manage, and scale your AI agents:

  • Embedded LanceDB Memory (RAG 5.21): Replaces standard sliding-window context with a high-fidelity, columnar vector database featuring Native Hybrid Search (Ollama semantics + Tantivy FTS) and an Entity Intersection Reranker.
  • Background Markdown Generation: A deterministic Python script strips terminal noise, reducing context weight by 85%. A background daemon then weaves this into a human-readable Markdown wiki (memory-wiki/).
  • GitOps Enforcement: AI agents are forbidden from coding on main. They must create GitHub issues (aim bug), branch out into isolated worktrees (aim fix), use TDD, and deploy atomically (aim push).
  • Interactive TUI Cockpit: A visual terminal interface (aim tui) to configure LLM routing, guardrails, and context limits without editing JSON files.
  • Cognitive Routing: Route expensive coding tasks to flagship models (e.g., Gemini Pro) in your terminal, while offloading repetitive background tasks (like memory indexing) to free, local models (e.g., Ollama) on your GPU.
  • P2P Knowledge Cartridges: Package thousands of pages of documentation into pre-vectorized Native Apache Arrow .parquet files. Share and download them peer-to-peer via BitTorrent (aim export / aim jack-in) to give agents instant recall of entire frameworks without burning API tokens.
  • Universal IDE Support (MCP): A built-in FastMCP server exposes the memory databases to any connected IDE (Cursor, VS Code, Claude Desktop) without requiring platform-specific adapters.
  • Crash Recovery & Handoffs: When the context window fills up, run aim reincarnate to extract active context and spawn a fresh terminal session. If the CLI crashes, run aim crash to salvage the interrupted session.
  • Anti-Drift Shield: A background hook continuously tracks autonomous tool calls. Every 50 actions, it forcefully halts execution and requires the agent to recite its GitOps rules, preventing "Lost in the Middle" context degradation.
  • Peer-to-Peer Wiki Sync (Syncthing): Offload heavy memory compilation to a secondary server by syncing the memory-wiki/ folder natively via Syncthing.

📖 Essential Documentation

To get started with A.I.M., follow the guides below. For deep architectural lore, benchmarks, and philosophy, visit the A.I.M. Knowledge Base.

☕ Support the continuous evolution of A.I.M.: Buy Me a Coffee

Navigation

Deep Dives & Philosophy

👉 A.I.M. Knowledge Base (Explore the full design history, benchmarks, and architectural maps in our public Obsidian vault).

Clone this wiki locally