Put Amorfs against a real problem

AI agents are good at writing, coding and searching text. They still struggle when answers have to be exact — complex relationships, changes over time, precise attribution. Amorfs turns massive, complex datasets into structured, source-linked knowledge an agent can query directly — without flooding the context window or requiring live database connections.

On prepared packs we have run — including a synthetic refinery unit of about 258,000 facts — an agent can ask for a named record and get a small answer back, typically a few hundred tokens, without a live database connection. We have not published an independent benchmark against retrieval. The scorecard is how we find out on the partner’s twenty questions.

The missing piece

Companies want agents that can answer from their systems-of-record: which asset is installed, who controls an entity, what was true on a given date, and which source supports the answer.

Connecting agents to live systems gives them too much power. Pasting extracts into chat blows context windows. Making models write queries is wrong too often to be reliable.

A Knowledge Pack is a versioned, immutable snapshot. No query language for the model to invent, no embedding service, and no live connection into production data. Nothing leaves the client, and no third-party cloud services are needed.

Prove it without touching production. No production credentials, no cloud ingest, no new hosted data platform — the extract stays in the client environment.

Most connectors give access. They do not give knowledge.

• App connectors such as Gmail or Drive read the user’s own world. They do not hold reference knowledge.
• API wrappers grow a tool for every endpoint. Tool choice is where agents make mistakes.
• Search returns relevant passages with no completeness guarantee.
• Databases have full query power, but the schema does not explain itself.

A Knowledge Pack is different. Every pack uses the same small set of tools, so adding new knowledge doesn't add new tools. The pack describes itself: what it holds, how records link, and starter queries that show how and where to find records. Records come back structured, with time and source on every link. If anything was cut, the answer says so.

That is why a pack complements Gmail or Drive rather than replacing them.

How it works

Three simple steps: open the pack, read the map, then describe the required record. Follow-up questions are just edits, not new searches.

1. Open the pack. The pack is a local file or folder that the agent can read — not a live connection to a database.
2. Read the map. The agent learns what kinds of records the pack contains. One read is enough: record counts, how values are distributed, samples, and how records link. The agent does not need prior training on the dataset.
3. Ask for a record. Describe the record you want back. The pack fills it in. The next question is an edit of the same record, not a new search.

Packs sit beside the tools you already use. Markdown and document bundles remain the right fit for docs and wikis. Retrieval remains the right fit for finding relevant text — or choosing which pack to open. Packs are for the bounded, structured records inside the selected domain.

The evaluation

A four-week evaluation: one bounded domain, one masked extract versioned into a Knowledge Pack on your machine, and 20 real questions with known answers, measured against your current approach. This is a paid evaluation.

• One client and one bounded domain — a single process unit, legal matter, ownership structure, or product family.
• One masked or anonymised extract, versioned into a Knowledge Pack on your machine. The extract does not need to leave that environment.
• Twenty real questions with known answers, agreed before scoring.
• A comparison with the current approach — retrieval, generated SQL or graph queries, or a person using the source system.
• A report scorecard: exact correctness, response size, whether the same question against the same pack still returns the same answer later, and evidence quality — can the agent show the fact it used, where it came from, and when it was valid?

First Cognition supplies the knowledge infrastructure, modelling, pack pipeline, agent integration and scorecard. The partner supplies the domain, client relationship, masked extract, subject-matter input and comparison baseline.

If you lead a plant or asset data team, MDM or data governance, compliance or financial crime, legal operations, or another agent programme on connected records, start with a twenty-minute demo on the synthetic unit.

The first step is a 20-minute live demonstration using synthetic refinery data. If there is a fit, we agree the bounded domain, evaluation fee and four-week schedule before any client data is used.

Do not send extracts to this website. If we proceed, the extract is versioned on your machine under a signed evaluation agreement.