Public leaderboards score answers to fixed questions. Pensiv ships a different answer shape — a Decision-Grade Brief, not a ranked list — so leaderboard rank tests the wrong thing. What we run instead, and why every claim has to trace to source, is below.
Three expert raters, blind to our marks, scored the economic case. Every number below traces to the eight-document study, available on request.
We don't publish a leaderboard score — those get walked back. We publish numbers an independent rater can check. See how we tested →
Useful memory isn't a passive log. It's an active structure — recurring roles, actions, and constraints — that gets rebuilt as new evidence comes in. A search system that just returns a ranked list of passages drops that rebuilding work back on you.
Pensiv works out its own structure as material builds up. By the time you ask, the answer is shaped by everything your team's knowledge has seen — not just by what shares words with your question.
The valuable cross-team connection — security's "lateral movement" and SRE's "cascading failure" — has no shared words. Keyword search misses it. Meaning-based search hedges on it. To find it reliably, the system has to compare the underlying pattern, not the exact words.
Pensiv finds the prior case even when it doesn't share a single word with your query. Legal precedent, competitor moves, clinical patterns — matched by the underlying pattern.
Memory has to age. Neither "store everything forever" nor "delete after N days" matches how knowledge actually fades in practice. What matters is what people keep coming back to, what gets replaced, and what stops being relevant — without losing the trail of what came before.
Important knowledge sticks around. Irrelevant knowledge fades. When new information contradicts old, the old one is linked to its replacement, not overwritten. The system forgets by ranking things lower, not by erasing them — every earlier state stays on the trail.
Twenty-seven-persona Monte Carlo. Dual-baseline comparison against running with no memory at all andagainst the commodity AI-memory tier. Failure-mode catalogue scored by three independent expert raters blind to the author's marks. Twelve kill criteria you can hold us to in a thirty-day pilot.
$950,000 NPV per mid-ACV deployment at a 12% discount rate over three years · 2.6-month payback at the median customer · 56.7% process-risk reduction (median across three independent raters, range 7.9%–76.2%). Every number traces to the canonical eight-document study, available on request.
Security writes "lateral movement." SRE writes "cascading failure." HR writes "escalation pattern." Same underlying problem, five different vocabularies, no shared words. Keyword search has no way to connect them. Meaning-based search hedges and hands you the closest noun, not the closest pattern.
Pensiv matches by the underlying pattern. The same pattern shows up across all five teams — without anyone having to agree on a single term for it.
What we measured: same-pattern pairs from a test set drawn across different fields show up in the top results, even when they share no keywords. Open the product, click any claim it puts together, and walk it back to the original source.
Evals score answers to fixed questions. We're shipping a different kind of answer, so leaderboard rank isn't the right test. What we run instead — and what each gate measures — is below.
Defense tabletop exercises, partner-handoff simulations on legal matters, multi-year lab data rolled into one view for clinical. Measured in hours-to-get-up-to-speed, patterns recovered, decisions reversed.
Every claim the brief makes traces back to its source. The audit trail is the proof — open the product, click any claim, and walk it back to the original record.
Every feature is held back by a tiered test suite covering correctness, contracts, and stability. A feature only ships when its tests pass. The number of tests grows with the product — what matters is whether the tests are passing, not how many there are.
The source trail isn't a logging layer bolted on. Every search returns the path that produced the result. Every claim the system puts together links back to a source. In regulated industries that's the price of admission, not a feature.
Three independent research programs reached the same conclusion: curated knowledge turns a smaller, cheaper model into a better one. The knowledge layer — memory that fades by importance, same-pattern matching, automatic pattern discovery — is the differentiator. Not the model size.
Reproducible across model families| Capability | Pensiv | A typical AI search setup |
|---|---|---|
| Source trail on every answer | Yes | Rare |
| Confidence breakdown by component | Yes | No |
| Temporal accuracy (when did we know what) | Yes | No |
| Immutable write trail for audit | Yes | No |
| Decision-Grade Brief output (not just retrieval) | Yes | No |
A small AI with good knowledge beats a much bigger one without it — for a fraction of the running cost. The model isn't the thing holding you back. Memory is.
