Skip to main content

Grounding & Memory

InaiAgents don't answer from thin air — they retrieve and reason over your real project data.

Retrieval, not prediction​

When an InaiAgent answers, it first retrieves the relevant project records, then reasons over them. The insight is anchored to what your data actually says.

What Mira grounds on​

Before Mira answers a portfolio question, she assembles the current state of several distinct registers — read fresh from your live data on every question, never from a cached summary. They are deliberately kept apart, because they answer different questions and conflating them produces confident nonsense.

Portfolio — the shape of the work. How many programs you have, total tracked budget and its capital/operating split, how programs are distributed across stewardship tiers, your largest programs by budget, and a roster of every open risk record with its severity and score.

Financials — the money, with its working shown. Each program's planned budget alongside its money-at-risk, the arithmetic that produced it, portfolio totals, and your real month-to-date AI compute spend with a month-end projection that is always labelled an estimate rather than presented as a figure.

Governance — what you're assessed against, and what's been raised. Every control in force, with its domain, whether it is gate-bearing or advisory, which programs it applies to, and the criteria text an assessment interprets. Alongside it, the findings — kept strictly distinct as Suggested (an AI proposal nobody has accepted), Open (human-confirmed), or Closed (resolved or dismissed). Mira will not describe a Suggested finding as a confirmed breach.

Review posture — how much of the portfolio has actually been looked at. Every program falls into exactly one of four cohorts: reviewed (a human has authored a risk assessment), awaiting sign-off (the AI flagged something no human has confirmed), not yet reviewed (work is overdue and no assessment exists at all), or nothing currently flagged. That last cohort is the one worth dwelling on: nothing flagged is not verified safe. Those programs carry no assessed severity because nobody has assessed them, and Mira is instructed to say so rather than let silence read as a clean bill of health.

The change record — what has actually moved. Every other register describes the current state, a snapshot. This one describes movement: which items changed, on which fields, from what to what, and when. It also carries its own horizon — the date the record begins — so Mira can never imply that nothing happened before it started watching.

One program in focus — when you ask about a specific program by name, Mira narrows to that program's governance rather than answering from the portfolio-wide picture. If nothing applies to it, she says so and names the two things she checked, so the negative answer is verifiable rather than merely asserted.

Registers carry conduct, not just content​

This is the part that is easy to miss and does most of the work.

Each register hands Mira more than facts. It hands her the rules for talking about those facts — what to abstain on, which two things must never be conflated, and what a silence does and does not mean. A few of the standing rules:

  • Nothing flagged is not verified safe — absence of a finding is not evidence of health.
  • "Needs sign-off" is a review state, not a policy — your governance controls may include one literally named for sign-off; that is a different thing from a program awaiting review, and the two must never be swapped.
  • Never answer "what changed" from current totals — the totals are real, but they describe now, not movement.
  • An AI suggestion is not an organizational position until a human authors it.

The effect is that Mira's caution is grounded in the data itself, not bolted on afterwards as a disclaimer.

Every fact carries its source​

Alongside each figure, Mira carries a short label naming where it came from — which register, and on what basis. When you see a number in an answer, you can trace it.

Derived numbers go further: they show their arithmetic instead of asserting a total. A money-at-risk figure arrives as the budget and the proportion that produced it, not as a bare number you have to take on faith. Estimates are labelled as estimates. A projection says it is a projection.

The purpose is auditability. A figure you can check is worth more than a figure that merely sounds authoritative.

Confidence travels with the data​

Different tools expose different levels of detail. InaiBridge records how trustworthy each signal is and carries that confidence through to the answer, so you always know how much to trust a given conclusion. A less-detailed source degrades gracefully to "uncertain" — never to a confident-but-fabricated claim.

Crucially, it degrades to unknown, never to zero. A missing estimate is not an estimate of zero. A program nobody has reviewed is not a program with no risk. Treating an absence as a value is how a reporting system quietly starts lying, so InaiBridge refuses to do it — even where a zero would make a chart look tidier.

Three kinds of memory​

A grounded answer can draw on more than the current snapshot:

  • Current state — what your connected records say right now (what's at risk, what's overdue).
  • Similar past situations — earlier projects that resemble the current one, especially ones that slipped, so a risk is judged against precedent rather than in isolation.
  • Recalled context — relevant things noted earlier (for example, a priority leadership has flagged) that bear on the current question.

A Project Intelligence briefing composes all three; a quick question may only need the first.

Two things worth knowing about how these behave in practice. Similar past situations can only draw on comparisons your history actually contains — a first-of-its-kind programme has no precedent to be judged against, and Mira will say so rather than reach for a loose analogy. And recalled context is deliberately explicit: it holds what has been put there to remember, not everything ever said. That is a design choice, not a limitation to route around — memory you didn't consent to is not a feature.

Ground or abstain​

When your data doesn't hold the answer, an InaiAgent says so — plainly — rather than inventing one. It would rather tell you "the records don't show that" than manufacture a confident-sounding claim. Grounding is the point: reason over what's real, abstain on what isn't.

That judgement is made deliberately early, and deliberately cautiously. Before answering, an InaiAgent decides whether a question needs your project data at all — and the decision is biased toward retrieving it. Only a clearly self-contained question (a definition, a greeting, arithmetic) skips the lookup; anything ambiguous grounds anyway. The asymmetry is intentional: an unnecessary retrieval costs a moment, while a wrongly skipped one produces exactly the ungrounded answer this product exists to prevent.

When two measures share a name​

Some questions have more than one honest answer, and picking one silently would be the easy mistake.

"How much money is at risk" is the clearest case. It can mean containment — the whole budget sitting inside programs that are overdue or unreviewed, because a troubled programme puts all of itself at stake — or expected exposure, each budget scaled by how much of that programme has actually slipped. Both are real. They are different numbers, and the smaller one is not a correction of the larger.

Mira gives you both and names the difference, rather than choosing one and appearing consistent. Containment is the headline; expected exposure is what ranks programs against each other.

When there is no honest answer​

Cost of delay is the example we'd rather not advertise, and the one that best shows the principle.

There is no figure InaiBridge trusts enough to state, so when you ask what delay is costing, Mira declines and says why. She does not estimate it, does not derive a plausible-looking proxy, and does not quietly answer a nearby question instead. An abstention that looks like a gap is more useful than a number that looks like an answer.

Mira grounds on what you're looking at​

Mira, the conversational companion, grounds on your data and on the surface you're using:

  • In the Portfolio, it can scope to the program you've selected.
  • In the Inbox, it grounds on the approval gate you're viewing.
  • In the Canvas, it draws on your project's context.

Mira always states what it's grounded in, so the scope of an answer is never ambiguous — and, like every InaiAgent, it answers only from what that scope actually contains.

What you're entitled to see​

Mira answers only from the data your role entitles you to see. A program outside your scope is absent from her grounding entirely — she is not filtering an answer after forming it, and there is no fuller version of the answer being withheld. She reports on the portfolio as you can see it.

One consequence is worth stating plainly, because it affects how you read every count in an answer: figures like "15 programs not yet reviewed" describe the programs visible to you, not necessarily your whole organization. Mira qualifies counts this way in her answers for exactly that reason.

Where this sits​

This page is about what Mira knows and why you can trust it. For how you work with her — scoping to a selection, chart lenses, voice input, spoken replies, and language selection — see Ask about what you're looking at in the Portfolio guide. The Portfolio's Honest confidence section covers how the same discipline shows up in that app's own surfaces.

One point of vocabulary worth separating, since both are called drift. The Portfolio's Automated Drift Detection is a sweep: it watches for meaningful movement and proposes risks for a human to confirm at a gate. The change record described above is the evidence layer underneath — the raw record of what moved, which Mira reads to answer "what changed recently?" The sweep acts on movement; the change record simply remembers it.