Synthetic research is only worth anything if you can defend it. Here is exactly how a persona is built, how a report is assembled, and where the limits are.
The premise
“We do not react to reality, but to the models of reality in our heads.”
— Daniel Kahneman, Nobel Laureate in Economics
Objective
Conversion rates, click-throughs, bounce rates. Numbers without context for why someone actually acted.
Subjective
Motivations, hesitations, and the reasoning behind a reaction. This is what a structured interview surfaces that a dashboard can't.
How a persona remembers
Flat retrieval can surface a relevant-sounding paragraph with no guarantee that a stale, contradicted, or one-off statement isn't carrying the same weight as a core trait. Persona memory is structured to prevent that.
Ask the persona
Planner selected: Preferences · Experiences — the rest is never retrieved.
Facts are extracted into biography, experiences, preferences, social circle, work, and psychometrics — and a planner selects the domains relevant to each question instead of dumping a whole profile into the prompt.
Every fact carries a reference count and a last-referenced timestamp. A heat score promotes well-corroborated facts from short-term to long-term, so what shapes an answer is what's actually current about that person.
Facts link to related facts on typed, weighted edges, and unreferenced ones decay on a forgetting curve. Core personality traits sit in a decay-exempt tier, so a persona's fundamental character doesn't erode just because it wasn't queried recently.
The model synthesises three published architectures — six-domain extraction (Synthius-Mem), time-tiered freshness (Memory OS), and graph connection with decay (Mnemosyne) — on self-hosted pgvector. No single open-source system combines all three; that combination is the actual differentiation, not the use of retrieval itself.
What every response captures
One response
I've used the jar for years, but honestly the refill works out cheaper per use — I'd try it once if the pack showed the maths.
Same answer, four separate signals — extracted independently, never averaged into one score.
The persona's own words, stored against the question that produced them.
How a report gets built
Every report is assembled deterministically from structured stage output. Nothing is generated in a single free-form pass, because a single pass can't be inspected.
Reads the approved questions and panel composition, then plans what the report has to cover.
Pulls recurring themes out of every transcript, each carrying the question that produced it.
Run in parallel: one synthesises findings across the panel, the other scores purchase intent first and anchors sentiment to it.
Turns themes and scored sentiment into recommendations tied to the decision you asked about.
Assembles every section deterministically from the previous stages' structured output.
Checks every claim in the executive summary against the real pipeline data. What can't be traced doesn't ship.
Before a report saves, the Validator extracts every numeric and thematic claim in the executive summary and checks it against the actual interview data. Every check is logged, matched or not, so a claim that can't be traced back to a real response gets stripped rather than shipped quietly. If a run comes back short of the panel you selected, the report says so on its face.
See the pipeline run on your own concept, and read the report it produces end to end.
Start a free simulation