VAAS-X Cognitive Memory MCP server

vaas-x

Persistent, outcome-grounded memory for AI agents and connected devices.

pip install vaas-x gives any device or agent a persistent, queryable record of what it's seen — what happened, what it did, what the outcome was — and retrieves the closest matching past episode in milliseconds, on CPU, with no GPU and no vector database to stand up yourself.

Install

pip install vaas-x

Get a free-tier API key at vaasx.com — no card required.

Quickstart

from vaasx import Bootstrap

brain = Bootstrap(api_key="...", device_id="my_device")

# Point it at a structured data stream (sensor readings, agent actions,
# API events -- anything JSON-shaped) and it profiles the stream
# automatically, no schema design required.
brain.connect("https://your-data-stream.com/events")

# Query for the most similar past episodes, ranked by what actually
# worked, not just what's nearest in vector space.
hits = brain.query("engine running hot", k=5, prefer_success=True)

An agent that already knows its own outcome the moment it acts can log a full state/action/outcome episode in one call instead of a separate ingest-then-tag-later round trip:

resp = brain.ingest([{
    "state": {"observation": observation},
        "action": {"taken": action},
            "outcome": {"success": result.success},
            }])
            episode_id = resp["episode_ids"][0]

            # Report back on a retrieved episode once you know how it went
            brain.outcome(hits[0]["id"], success=True, delta=0.12)

Three calls — ingest, query, outcome — cover the whole loop. No schema to design, no vector DB to host, no embedding model to serve yourself.

## What it does

schema classification, and anomaly detection all run locally before anything touches the network.

• Profiles a raw data stream automatically — statistical profiling,

weighted toward what worked last time in a similar situation, not just nearest by embedding distance.

• Stores episodes as (state, action, outcome) triples, so retrieval can be

no domain-specific build. Verified live on a £30 ESP32 microcontroller and on a standard EC2 instance.

• Runs on both a small edge device and a standard server — same codebase,

## Numbers, not adjectives

m7i-flex.large CPU instance — no GPU, no AVX-512.

• 14ms mean retrieval across 1.18M stored episodes on a standard AWS

given 24 sensor channels with every identifier stripped, correctly classified all 24 with zero prior domain knowledge.

• Blind-tested on NASA's CMAPSS FD001 predictive-maintenance benchmark:

subjects on a real wearable IMU dataset (UCI Daily and Sports Activities).

• Consistent zero-config classification result across 8 independent

Every number above is independently reproducible — see below.

## Reproduce the benchmarks yourself

Every result this SDK's marketing points to has a step-by-step guide against a real public dataset or environment, with your own API key, plus an independent local cross-check that doesn't depend on the SDK at all:

Shortened. The full README is on GitHub.