fajarhide/omni

Noise-canceling context and long-term memory for your AI agent. Stop paying Claude to read 10,000 lines of terminal noise like a headphone for AI agent

What it solves

OMNI is designed to stop AI coding assistants from wasting tokens and budget on "terminal noise"—the thousands of lines of irrelevant logs, progress bars, and boilerplate output generated by commands like npm install or docker build. It also addresses "AI amnesia," where agents lose project context, goals, and previously solved bugs when sessions are restarted or when switching between different AI IDEs.

How it works

OMNI acts as a local interceptor for terminal executions. It uses a deterministic pipeline (Read → Guard → Score → Collapse → Distill → Persist) to filter out noise and keep only the critical signal (e.g., errors and stack traces).

Key components include:

Content Distiller: Reduces output volume while preserving essential information.
RewindStore: A local SQLite database that stores the full, raw logs. If an AI agent needs the original noise, it can retrieve it using a hash.
Memory OS: A persistent layer that allows users to set long-term goals (omni goal) and save project-specific rules or fixes (omni remember) via semantic search.

Who it’s for

Developers using agentic AI coding tools such as Claude Code, Cursor, Windsurf, Roo Code, OpenAI Codex, or Antigravity CLI who want to reduce API costs, speed up response times, and maintain persistent context across sessions.

Highlights

Token Reduction: Can reduce bytes reaching the model by up to 85% in some cases, specifically on noisy tools like git and cargo.
Cross-Session Memory: Automatically injects session summaries and project goals into new agent sessions to prevent re-explaining the project.
** لقب MCP-compatible**: Built in Rust for high performance and low latency.
Fail-Open Design: If the tool panics, it passes raw output through to ensure the host agent never crashes.
Customizable: Users can add their own noise filters via TOML configuration files.