Each person’s agent builds a Cognee memory graph from their own messages or an imported ChatGPT/Claude export. Superlinked SIE Serverless runs every model task on small task-specific models: picking the personal conversations out of an import, extracting facts with a JSON schema, redacting PII from the public agent card, embedding, reranking, and running the agent-to-agent pre-date and its verdicts. Every fact keeps the exact quote it came from; untraceable facts are dropped.
SingleStore holds the agent directory: one SQL query combines hard filters (verified only, ages, genders, mutual preferences) with vector similarity, and also stores the audit trail. The top candidates are reranked, then the two agents talk for a few turns, recalling answers from Cognee. Verdict reasons must quote the other agent’s actual messages. When both agents recommend the match and both people opt in, a booking agent picks a venue from shared interests and a time both calendars allow, and writes a calendar invite (reservation simulated).
Statsig’s match_policy config tunes pool size, rerank cut, turns and minimum score, and the matching_strategy experiment compares vector-only matching with agent pre-dates. Amplitude tracks the funnel from memory import to booked date, plus every model call with model and latency. The audit timeline shows each search, rerank, agent message, memory lookup, verdict and booking. All personas are synthetic.