An AI that remembers your organisation.
Every assessment, agent, policy, use case, decision, and experiment across your Fronterio tenant feeds a nightly consolidation brain. It distils durable learned facts — preferences, patterns, risks, aspirations — and recommends the next move, quietly, every day.
Fact types the engine tracks per organisation — from risk signals to stated aspirations
Surfaces the recommendations appear on: consultant, dashboard, digest, Mission Control
K-anonymity floor for every cross-customer insight — identity never leaves your tenant
How the Memory Engine works
Every action your team takes on Fronterio writes an immutable, Zod-validated signal to an append-only bus — no raw user IDs, no free-text, bounded payload whitelist. A nightly cron at 05:00 UTC reads every signal since the last run for each of your organisations and feeds them through a two-stage consolidation.
Stage one is deterministic: five rule-pass detectors scan for patterns a spreadsheet could catch — policy-publish streaks, assessment-score drops, agent-deploy cadence, use-case ship rate, guide-completion rhythm. These produce candidate facts with evidence trails pinned to specific signal rows. Stage two is a frontier AI model, gated to Pro+ and only when new signals or a consultant-inferred fact justifies the call. The model looks at your bucket context + prior facts + recent signals and returns up to twelve additional candidates — subtle preferences, stated aspirations, decision contexts the rule-pass can't see.
Candidates merge into a durable memory with supersede semantics: same fact at higher confidence reinforces, different value supersedes. A canonical 55-key vocabulary prevents drift; anything off-vocabulary lands in an admin review queue instead of polluting your memory. Facts carry confidence bands, evidence signal IDs, and an expiry per fact type — blockers expire in 14 days, constraints in a year.
Recommendations compute from memory + current state + (Business+) cross-customer cohort patterns. A single recommendation can opt into four surfaces — your consultant prompt (top five), your dashboard home card (top three), your weekly digest email (high priority only), and the Mission Control dashboard (full list). One source of truth, four renderings, always in priority order.
Why memory, not just analytics
Without the Memory Engine
- Your AI consultant starts every conversation from scratch — it knows your score but not your preferences, history, or decisions
- Pattern detection happens in someone's head — churn risks, stalled initiatives, and capability gaps sit in memory until a meeting surfaces them
- Every dashboard widget answers "what happened" but nothing answers "what you should do next, grounded in what you've done"
- Peer benchmarks are stale PDFs; "what orgs like us typically do" lives in a consultant's LinkedIn DMs
With the Memory Engine
- The consultant cites learned facts by fact_type with confidence markers — no hallucination, no repeating the same question every conversation
- Risk signals surface the morning they qualify, with evidence linking back to the specific agent incidents or score drops that triggered them
- Every surface — dashboard, digest, Mission Control — reads from one recommendation set, ordered by priority, auto-dismissed when no longer relevant
- Cross-customer cohort insights (Business+) are bucket-only with k≥15 anonymity — useful patterns, zero identity leak
What ships with the Memory Engine
Learned facts per organisation
Preferences, patterns, relationships, constraints, commitments, risk signals, aspirations, capability gaps, blockers, momentum, cohort insights — twelve canonical fact types.
Rule-pass + AI consolidation
Deterministic detectors for every org; Pro+ adds nightly Sonnet pass with signal-threshold gating and Zod-validated output.
Template-driven recommendations
Declarative templates map fact types to recommendations with plan-tier priority boosts, CTA links, auto-dismissal when the underlying fact clears.
Four rendering surfaces
Consultant prompt injection, dashboard card, weekly digest section, Mission Control dashboard — same source, different priorities.
Privacy by design
Org-salted hashing, Zod-gated emit boundary, bounded meta whitelist, bucket-only cohort aggregation with k≥15 floor enforced at the database layer.
Opt-out honoured at every layer
cohort_contribution_enabled flip excludes an organisation from both cohort writes and reads — no data leaves your tenant without explicit consent.
How it lands in your tenant
Signals start flowing
From day one, every assessment, agent approval, policy publish, guide completion, decision, and experiment emits a Zod-validated signal to the bus.
First consolidation run
Day two at 05:00 UTC: the rule-pass finds its first patterns, seeds the first organisation_memory rows. Consultant starts citing them.
Recommendations appear
Within a week, the dashboard home card and Mission Control show your first recommendations. The digest includes high-priority items every Monday.
AI pass kicks in (Pro+)
When signal volume or consultant-inferred facts justify it, our AI model adds the subtle facts — preferences, aspirations, decision contexts detectors can't see.
“We stopped explaining our organisation to the AI every Monday. It just knew. Three weeks in, it flagged a governance drift we hadn't noticed ourselves.”
Available across Pro, Business, and Enterprise
Rule-pass + consultant injection ships to every plan. Pro+ unlocks the nightly AI consolidation pass and dashboard-card recommendations. Business+ gets Mission Control and predictive cohort insights. Enterprise adds cross-customer risk-signal predictions and Brain API access.
Give your AI a memory.
Book a walkthrough and see what the engine would learn about your organisation in its first week.