Sage Memory

SAGE Memory — Give Your AI a Brain That Persists

SAGE (Sovereign Agent Governed Experience) is an open-source institutional memory layer for AI agents. Every memory goes through BFT consensus validation before it's committed — not a flat file, not a vector database with a search bar. Governed knowledge.

What It Does

• Persistent memory across sessions — your AI remembers what worked, what didn't, and why
• 4 application validators — Sentinel, Dedup, Quality, Consistency. 3/4 must accept (BFT quorum)
• 13 MCP tools — sage_inception, sage_turn, sage_remember, sage_recall, sage_reflect, sage_forget, sage_list, sage_status, sage_timeline, sage_task, sage_backlog, sage_register, sage_red_pill
• Works with any AI — Claude, ChatGPT, Gemini, DeepSeek, or any MCP-compatible model
• Runs locally — no cloud, no API keys, your memories stay on your machine
• Real consensus — CometBFT under the hood, Ed25519 signed transactions, deterministic validator keys
• CEREBRUM Dashboard — real-time brain visualization, domain stats, memory detail, focus mode

Quick Start

1. Download from GitHub Releases (macOS DMG, Windows installer, Linux tarball)
2. Run the setup wizard
3. Add the MCP config to your AI tool
4. Call sage_inception — your AI wakes up with persistent memory

MCP Configuration

{
  "mcpServers": {
    "sage": {
      "command": "sage-gui",
      "args": ["mcp"],
      "env": {}
    }
  }
}

Research

Backed by 4 published papers. Key result: agents with SAGE memory achieve Spearman rho=0.716 learning improvement over sequential runs. Without memory: rho=0.040.

• Paper 1: Agent Memory Infrastructure (Byzantine-Resilient Institutional Memory)
• Paper 2: Consensus-Validated Memory Improves Agent Performance
• Paper 3: Institutional Memory as Organizational Knowledge
• Paper 4: Longitudinal Learning in Governed Multi-Agent Systems

Links

• GitHub: https://github.com/l33tdawg/sage
• Landing Page: https://l33tdawg.github.io/sage
• Author: Dhillon Andrew Kannabhiran (@l33tdawg)
• License: Apache 2.0