01 · QA accuracy
92.8%
End-to-end question answering across all six question categories, LLM judged with the benchmark's official prompts. Above published SOTA (Supermemory at 81.6%).
+11.2 vs SOTA
FURO turns the things agents read, hear, and write into typed records your team — and your models — can query by field, not by keyword.
Every call, email, doc, or tool output lands in FURO and crystallizes into a lattice of typed records — companies, people, deals, parts, orders. No prompt tricks. No per-app glue code.
Transcripts, tickets, call notes — whatever comes in is parsed into entities, relations, and fields the moment it lands. No re-ingestion, no nightly jobs.
Units normalize. Dates snap to ISO. Entities resolve against your canonical lists. By the time your agent queries, there's one right answer for the same fact.
The same customer, part, or record links across CRM notes, ERP orders, field-service jobs, marketing campaigns. Query the node — get everywhere it's been.
Below, the same question runs two ways. Vector search surfaces keyword-adjacent noise from unrelated records. FURO dereferences the typed field on the right record — every time, with the right scope.
Typed memory is aggregateable. Ask for rollups, filters, pivots, counts — the same way you'd ask a person who knows your business cold. Get a number, not a list of links.
Vector search caps out at "top-k chunks." Structured memory rolls, pivots, and compares across thousands of records — with the same interface a person uses.
Rolling soil pH? ARR booked this week? Open blockers by stage? FURO returns a value — typed, cited, sourced — not a list of loose links for you to sift.
The substrate doesn't care whether the record is a sales deal, a soil sample, or a repair ticket.
The shape does.
LongMemEval-S is the standard for long-term conversational memory: 500 questions spanning single-session recall, cross-session aggregation, temporal reasoning, preference retrieval, and knowledge updates. We run end-to-end ingestion, retrieval, and answer generation, graded by LLM-as-judge using the benchmark's official prompts.
| System | Overall | SSU | SSA | SSP | KU | TR | MS | R@5 |
|---|---|---|---|---|---|---|---|---|
| Full context (gpt-4o) | 60.2 | 81.4 | 94.6 | 20.0 | 78.2 | 45.1 | 44.3 | — |
| Zep | 71.2 | 92.9 | 80.4 | 56.7 | 83.3 | 62.4 | 57.9 | — |
| Supermemory | 81.6 | 97.1 | 96.4 | 70.0 | 88.5 | 76.7 | 71.4 | — |
| MemPalace (raw) | — | — | — | — | — | — | — | 96.6 |
| MemPalace (hybrid v4) | — | — | — | — | — | — | — | 98.4 |
| FURO | 92.8 | 96.9 | 100 | 100 | 90.0 | 90.0 | 90.0 | 99.1 |
Point FURO at your domain and it shapes itself around your vocabulary. The sales team gets deals, attendees, competitors. Agronomy gets blocks, samples, treatments. Same substrate, different schema.
A 32-minute call becomes a typed deal card — prospect, stage, close date, owner, blockers, competitors, attendees — all addressable, all queryable, all pivotable.
We don't ship industry SKUs. We ship a memory system. Your schema — tables, fields, relations — is how FURO knows what your business looks like.
No infra to stand up, no vector store to babysit. Connect your sources, describe your shape, and your agents have structured memory the next time they run.
Define the entities and fields your business runs on. YAML, UI, or inferred from samples.
5 minCalls, emails, docs, tool outputs, Slack threads. Ingest flows into the shape.
one-clickA typed memory tool your agent calls by field. No RAG rewrites, no chunking strategy to babysit.
3 linesPeople and agents query the same substrate. Answers cite their source — per-field.
live