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
- Profiles a raw data stream automatically — statistical profiling,
schema classification, and anomaly detection all run locally before
anything touches the network.
- Stores episodes as (state, action, outcome) triples, so retrieval can be
weighted toward what worked last time in a similar situation, not just
nearest by embedding distance.
- Runs on both a small edge device and a standard server — same codebase,
no domain-specific build. Verified live on a £30 ESP32 microcontroller
and on a standard EC2 instance.
## Numbers, not adjectives
- 14ms mean retrieval across 1.18M stored episodes on a standard AWS
m7i-flex.large CPU instance — no GPU, no AVX-512.
- Blind-tested on NASA's CMAPSS FD001 predictive-maintenance benchmark:
given 24 sensor channels with every identifier stripped, correctly
classified all 24 with zero prior domain knowledge.
- Consistent zero-config classification result across 8 independent
subjects on a real wearable IMU dataset (UCI Daily and Sports
Activities).
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:
- [CMAPSS predictive maintenance](https://vaasx.com/guides/cmapss.html)
- [Room-occupancy detection](https://vaasx.com/guides/occupancy.html)
- [Memory-augmented CartPole](https://vaasx.com/guides/cartpole.html) (zero download)
- [Wearable activity classification](https://vaasx.com/guides/sports_wearable.html)
- [Physics simulations](https://vaasx.com/guides/scientific_computing.html) (QEC, lattice field theory, Ising model)
- [Live drone control loop](https://vaasx.com/guides/drone_control.html)
Full guide index: [vaasx.com/guides](https://vaasx.com/guides)
## Tiers
Free gets you an API key instantly. Developer (£19/mo) and Professional
(£99/mo) are self-serve paid tiers with higher episode limits and
additional capabilities. Enterprise is contract-based, for teams wanting
on-prem or air-gapped deployment. See
[vaasx.com/pricing](https://vaasx.com/pricing) for current details.
## The honest caveat
This package ships as a compiled wheel, not source-available — the
retrieval and indexing internals are proprietary and protected as a trade
secret. What *is* open: full API documentation, working examples, and the
wire format, so you can see exactly what goes in and comes back even
without reading the internals.
Multimodal support (vision/audio/depth/thermal, via the optional `ac`
extra) is real but still early — encoder models are in active training
and not yet production-tuned; treat it as a preview, not a finished
feature.
## Issues & questions
This repo doesn't carry the source (see above), but it's the right place
to file bugs against the published package, ask integration questions, or
request a reproduction guide for a dataset/environment not covered above.
## Links
- [Documentation](https://vaasx.com/docs)
- [Whitepapers](https://vaasx.com/whitepapers) — technical deep dives, one per use case
- [Homepage](https://vaasx.com)
## License
Proprietary — see [LICENSE](LICENSE).