Vestige

Cognitive Deterministic Memory Transaction-Security OS for Agentic AI.

Your agent can think anything. Vestige decides what it is allowed to do.

It remembers every decision your project ever made, reaches backward through time to find the quiet choice behind today's failure, and blocks the actions your agent should never take. Receipt, one-use permit, effect. Or STOP with zero effects. Local. Encrypted. No cloud, no telemetry.

Install · Why not RAG · The Live Gate · Continuity · Benchmark · Science · Docs

The cause never looks like the bug

Agents re-learn the same lessons. They recommend a change you already tested and rejected, re-derive a fix that was already written down, and treat every session as if the last one never happened. Vestige is the deterministic memory security OS that ends that. Any MCP-capable agent writes memories as you work and retrieves them later: redundant memories merge, contradicted ones are flagged, unused ones fade, and when a failure hits, Vestige reaches backward to the decision that set it up.

_{A labeled fixture store, a real run. A SIGSEGV on startup in an arm64 container, and the version pin set 23 days earlier that shares zero words with the failure. Similarity ranked the pin fourth. Backfill ranked it first. It names the suspects and never calls the verdict. Watch the 58 second walk.}

Install

You need Node.js. No Docker, no signup, no compile step. Prebuilt for macOS ARM and Intel, Linux x86_64 and arm64, Windows x86_64.

npm install -g vestige-mcp-server@latest

Prefer Homebrew?

brew install samvallad33/tap/vestige

Scripted installs work too: eget samvallad33/vestige pulls the right prebuilt binary from Releases.

Connect it to your agent. Every MCP client understands this config:

{
  "mcpServers": {
    "vestige": { "command": "vestige-mcp" }
  }
}

Verify it: vestige dashboard, then open http://localhost:3927/dashboard. First run downloads a 130MB embedding model once; after that Vestige is fully offline, forever. Full walkthrough: docs/GETTING-STARTED.md.

Why not just RAG?

RAG retrieves text that resembles the query. That works when the answer looks like the question. It fails when the cause of a problem looks nothing like the symptom.