Phoenix Book a drill

Institutional memory for regulated organisations

Your organisation's memory, under your control.

Phoenix is a team of AI specialists that works on your own servers. What it learns is kept in your database, shared across the whole organisation, and can be traced and taken back.

Run the recall drill How the memory is shared
  • Your database
  • Your hosts
  • The model you choose
A house holding one shared memory store. Four specialists inside are connected to it. A cloud floats outside.

People leave. Models change. The memory stays.

Every day your teams teach AI how the organisation works: what was decided, which rules apply, what went wrong last time. With a cloud assistant, that knowledge builds up at the vendor.

With Phoenix it builds up in your own database, as ordinary Postgres tables. Replace the model and nothing is lost. Replace the vendor and you still have it.

  • Decisions
  • Policies
  • Lessons
  • Skills
  • Glossary
  • Roles
  • Project register
Someone leaves
The model changes
The memory stays

How the memory grows

Fed by your sources. Built up by your work.

Phoenix learns from what you give it and from the work your teams do with it. It turns that into lasting knowledge slowly and on purpose, and it keeps the source attached at every step.

Where it comes from

  • Your documents

    SOPs, technical files, runbooks, PDFs and spreadsheets. Each one is registered as a source, with its version, before anything is learned from it.

  • Official publications

    EU regulation and guidance, ingested with their version. When a text is replaced, the old version is marked as superseded.

  • Daily work

    Questions, answers, plans and decisions from your teams' work with the specialists. Each decision is recorded with the sources it rests on.

Everything that comes in is read as information, never as instructions. A document can inform an answer. It cannot change the rules.

How it compounds

Four layers stacked like steps. Notes at the bottom, then lessons and decisions, then judgment, and the signed-off canon at the top. Notes Lessons and decisions Judgment Canon
  1. NotesWhat is used and confirmed in daily work is kept. Themes that keep coming back move from the working notebook into lasting memory.
  2. Lessons and decisionsOvernight, Phoenix removes duplicates, rescores what it holds, retires what has gone stale, and records lessons and decisions.
  3. JudgmentPatterns that are confirmed again and again are promoted into the specialist's judgment, through six checks. Each promotion can be reviewed as a change before it counts.
  4. CanonWhat the whole organisation should know is nominated, signed off by a named person, and shared with every specialist at every site.

A new specialist starts on day one with everything your organisation has signed off. And because every layer keeps its sources, a recall reaches all of it.

The organisation canon

One memory for the whole organisation.

The shared memory belongs to your organisation, not to one assistant. Every Phoenix specialist reads it from the first question, at every site, with nothing to set up.

One signed organisation memory at the top, synced over locked links to teams at three sites. Each team has its own copy in the middle and four specialists around it. Organisation memory Head office Production Research lab

  • Shared by every specialist

    Decisions, policies, the glossary, roles and the project register reach every specialist automatically. Nobody opts in, and nobody is left out.

  • Private where it should be

    Each specialist's own working notes stay walled off by the database itself. Only signed-off knowledge is shared.

  • The same at every site

    Teams in different locations load the same memory. Each version is signed at the source, and every link between sites is mutually authenticated.

  • Keeps working when the line is down

    A site that loses its connection keeps using its local copy and catches up when the link returns.

How knowledge gets in

  1. NominateA specialist or a colleague points to something worth keeping.
  2. ReviewThe curator checks it for duplicates, conflicts and sources.
  3. Sign offA named person approves it. The approval is logged with their name.
  4. PublishIt becomes a new version. Nothing is edited in place, and every answer records the version it read.
  5. SyncEvery team at every site loads the same version.
Example entry signed
Kind
Policy
Title
Supplier audits follow SOP-QA-114, revision F
Classification
Internal
Signed off by
Head of Quality
Version
1284 · supersedes 1203
Sources
3 documents, 2 decisions
Reaches
Every specialist, 3 sites

Digital sovereignty

Sovereignty starts with who holds the memory.

Where the AI runs is half the question. The other half is where the knowledge it builds about you ends up. With Phoenix, it ends up with you.

Your hosts

Phoenix runs on your servers, or on a box we supply and you keep.

Your database

The memory is plain Postgres tables. Inspect them, back them up, take them with you.

Your model

Use any model, including open-weight models on your own hardware. Change it later and the memory stays.

Your budget

Priced per team per year. No token meter and no seat count.

If you choose a hosted model, it sees the passages used for each answer. Choose a local model and those stay inside too. The memory itself never leaves your database.

A server with the memory store built into its front. The AI model on top can be swapped between a local and a hosted one. The memory stays in the server. local hosted Your hardware, your database

Trace and recall

Your quality system can recall a faulty part. Can your AI recall a faulty source?

In Phoenix, every stored conclusion carries its sources. When one turns out to be wrong, you take it back in one step.

  • Traceable

    Every stored conclusion carries its sources, so you can see where any answer came from.

  • Retractable

    Revoke a source and everything derived from it is quarantined in one operation: memories, lessons, skills, decisions and canon entries. The audit trail stays intact.

  • Gated

    Actions without live evidence are refused. Claims it could not verify are disclosed in the answer.

Try itExample: classification guidance that was later amended

SourceGuidance note on app classification, 2019 editionActive
Derived from it
  • MemoryOur triage app is Class IActive
  • LessonSmall UI changes need no new reviewActive
  • SkillDraft section 4 of the technical fileActive
  • DecisionKeep the triage app in Class IActive
  • CanonSoftware classification policy, v12Shared
Relies on that decision
ActionSend the yearly classification summaryReady

Nothing revoked. Five items and one pending action rest on this source.

Keep Copilot or Gemini for everyday work.

Use them for mail, meetings and documents. Keep the knowledge you must own, and answer for, in Phoenix.

Three questions for any AI vendor

  1. Where does what the agent learns about us live, and can we take it with us?

  2. A source it learned from is wrong. Show us everything derived from it, decisions included, and quarantine it in one step.

  3. Can it run on our hosts, on a model we choose?

Ask us the same three. We answer them on your own data.

Made for teams that answer to a regulator.

Start here

Medtech and life science

Regulatory and quality teams. Documents too sensitive for public AI, and guidance that keeps being amended. A team of 10 to 30 people fits on one box.

With a delivery partner

Health data

Data and BI teams working with regional health data, where the data must stay inside the region's own walls.

Design partners

Legal and public sector

Firms and authorities that need a clear wall between matters and a way to take back what a wrong source taught. Talk to us about a design partnership.

Getting started

Start with a pilot that has to pass a test.

A paid, fixed-scope pilot on your own data. It passes when the recall drill passes.

  1. Install

    One team of specialists on your hardware, or on a box we supply. Your documents plus public EU regulation.

    10–30 people

  2. Work

    Real work on real questions. The shared memory builds up in your database as you go.

    8–12 weeks

  3. Recall drill

    Plant a wrong source, trace it, revoke it, show the quarantine, show a dependent action refused.

    Pass or fail

  4. Subscribe

    An annual subscription per team. Add teams and sites when you are ready. The pilot fee counts towards year one.

    Per team, per year

Run the recall drill on your own data.

hibsen@noetic-os.com
Write an email