How FRONTFACE works
Every decision a person or an AI agent makes draws on three things: what has already been learned, what things mean, and who is allowed to know them. FRONTFACE models all three once, so every human and every agent works from the same foundation.
The model
Ask “How is Acme doing this quarter?” and three questions must be answered before any number is worth reading.
Memory
Every session starts cold. Last month’s decision is made again, or lost.
FRONTFACE keeps decisions, preferences, and procedures, with where they came from and when they were true.
Meaning
“Acme” is three records in three systems. Each tool guesses which one you mean.
FRONTFACE resolves Acme once and links it to its tables, metrics, and documents.
Permission
The rule is a sentence in a prompt. The model may follow it.
FRONTFACE signs one decision about access and applies it in code, everywhere.
Most AI tools answer none of the three. They guess the meaning, forget the history, and leave permission to the model. FRONTFACE answers all three, then gives the model exactly what it needs: the right history, the right entity, and only what this person may see.
Memory
FRONTFACE turns the work you already do into memory: decisions, preferences, conventions, and procedures. It carries them into the next conversation and the next tool, so work builds on itself instead of starting over.
FRONTFACE reads what you already do: chats, coding sessions, tool calls. Everything lands in a ledger you own, on your machine.
A model you choose turns raw activity into short, typed memories. Each one is stamped with its source, when it was true, and who may see it.
Every turn starts with a short profile of what matters, read without a model call. Deeper search runs on demand. The original transcript is one step further, when you ask for it.
A procedure that repeats and succeeds is proposed as a skill. You read it, edit it, and approve it before it is saved.
“The Q3 customer review reports net revenue, not gross.”
Every memory carries two dates: when it became true, and when FRONTFACE learned it. That is how it can answer both “what was true in March?” and “what did we believe in March?”
| Kind | What it holds | Example |
|---|---|---|
| Facts | Exact, stable attributes | Timezone: Pacific. Default currency: USD. |
| Notes | Preferences in plain language | Prefers summaries that lead with the number. |
| Relationships | What connects to what, and decisions over time | The Q3 review reports net revenue. |
| Procedures | A pointer to a skill and its success record | Monthly close checklist: used 9 times, succeeded 8. |
| Style | The voice to use and the things to avoid | Plain language. No jargon in client emails. |
About you: your style, your preferences, how you work. It follows you.
About the work: decisions, conventions, lessons. It stays with the project and the people on it. For code, it travels with the repository.
On a project, FRONTFACE reads both and joins them at the moment of recall, linked by a pointer and never copied into each other. Leave the project and your personal memory leaves with you. A memory learned from something private never moves into a shared pool on its own.
Meaning
FRONTFACE keeps a map of your business, built from your own systems: the tables, the things they describe, the metrics people track, and the documents that discuss them. Each thing is resolved once, and every answer reads the same map.
| Object | What it is | Example |
|---|---|---|
| Entity | One real thing, however many systems name it | Acme Corporation |
| Metric | A measure with one agreed definition | Net revenue: gross revenue minus refunds and credits |
| Term | Business vocabulary, mapped to data | “Active customer”: a customer with a paid invoice in the last 90 days |
| Join path | How two tables connect | customers → orders → invoices |
| Passage | A piece of a document, linked to the entities it mentions | The renewal clause in the Acme agreement |
Your schema and business ontology: tables, columns, domains, metrics with their definitions, glossary terms, and the join paths between tables.
Your canonical entities: the one Acme that three systems call by three names. A cache of identity, not a copy of your data.
Your documents as passages, indexed for retrieval and linked to the entities they mention.
In the enterprise tier, a fourth graph connects the three to your organization: people, teams, roles, and the applications each person has connected. It lets FRONTFACE decide access in terms your business already uses.
FRONTFACE reads the structure of your connected sources: tables, columns, types, keys, and samples.
Business meaning is added: domains, entity types, metrics, glossary terms, and join paths. Enrichment is proposed and reviewable.
When a question names something, FRONTFACE resolves it in order: an exact match, then a close match, then a check against the source itself. The result is kept and reused.
Every answer reads the same map. Correct a metric definition once, and every answer that uses it follows.
Acme Corporation · Customer
Three systems, three names, one entity. Resolved the first time someone asks, and reused every time after.
Permission
FRONTFACE decides what each person and each agent may see, signs that decision as an access pass, and applies it everywhere. Memory, meaning, and answers are all filtered by the same pass. Nothing recomputes the rules, and the model never decides them.
A person, or an agent acting for one, makes a request.
FRONTFACE resolves every grant in play and signs a pass for this request.
Each part checks the signature itself and applies the pass. The map filters what it returns. Answers carry the row filters and column masks into the query before it runs. Memory filters what it recalls.
Change a grant and the revision moves. The next query, and the next recall from memory, drop what is no longer allowed.
Access pass · Signed
One pass per request. It covers the sources in play right now, not the whole company.
Most systems can revoke access to a database. Few can revoke what an AI already remembered from it. Every memory records the source it came from. When access to that source is withdrawn, the memory drops out of recall on the next request.
A rule in a prompt asks the model to comply. A rule in the pass is applied by code, before the query runs.
Read the benchmarkHow they work together
An AI agent asks FRONTFACE to “prepare our Acme review using our last meeting, current revenue, and open support cases.”
FRONTFACE signs a pass for the agent, scoped to what the account manager it acts for may see.
FRONTFACE recalls the decisions from the last Acme review, including that it reports net revenue. Only memories the pass allows come back.
FRONTFACE resolves Acme across the CRM, billing, and support, and supplies the definition of net revenue and the join path to invoices.
FRONTFACE queries revenue and open cases where they live, with the pass compiled into the query, and reads the review documents.
The agent receives the brief with its sources, the query behind each number, and any gaps stated plainly.
Every step is recorded on one trace, so you can see what was used and why.
Chooses which tool or skill fits the request
Decides who may see what, and signs it
Drafts the query and writes the answer
Applies row filters and column masks before the query runs
Distills memories from your work, using a model you choose
Recalls memory with no model call, filtered by the pass
Proposes a new skill
Saves a skill only after a person approves it
The delegation is recorded.
A helper that reads untrusted content, such as a fetched web page, is cut off from private memory. Anything it tries to remember is quarantined.
They are designed to connect through standards such as MCP. The same pass applies to every request.
Cloud connectors and hosted models may process data outside your infrastructure. We review the complete data flow with your team before you choose a setup.
No. Retrieval finds passages that look like a question. FRONTFACE models what things are: entities, metrics, tables, and the paths between them. Documents are one of its three graphs, linked to the entities they mention.
No. FRONTFACE holds the meaning of your data: its structure, definitions, and entities. Data is queried where it lives. Document passages and memory have their own stores, separate from your source systems.
Chat history keeps everything that was said. Memory keeps what was worth keeping: short, typed, stamped with its source and dates, and checked against your access every time it is recalled.
Yes. The ledger is a plain file on your machine, and the store is a database you can open. Correct a memory and the old version is superseded, not lost.
A prompt is an instruction the model may or may not follow. The pass is applied by code to the query itself. Our benchmark measures the difference.
The pass revision moves. The next query and the next recall drop what the person may no longer see, including memories that came from that source.
No. Someone working only with their own data has nothing to gate. The moment shared data is involved, the pass is required and every gate defaults to deny.
Which tool to use, how to write the query, and how to phrase the answer. It never decides access, and it never saves a skill without a person’s approval.
We will show you the memory, the meaning, and the permission behind the answer.
Examples are illustrative. They show the designed behavior, not a customer’s data.