The whole record in one scroll: the claims, the results with their scope, the figures, the economics, and the places the system fails. Every number here is one row of the ledger.
The AI stack today has two components: the model and the harness. We build the third - one with its own state and its own learning rules. The model stays frozen and swappable; the system around it learns while it works.
AI today generates. What it cannot yet do is connect. Sapience is the connective layer: one living memory that connects your sessions to each other, your AIs to each other, you to your AIs - and, with consent, you to other people.
Most network products are worth nothing until the network shows up. Sapience is at its most useful on day one, alone - the memory and long-context reasoning now carry public leaderboard receipts. The network isn't the price of admission; it's the compounding on top.
Every number, with its n, seeds, caveats and result path, is one row of the benchmark ledger. Nothing on this page appears without a row there.
| Benchmark | Our number | Comparison | Scope | Verified |
|---|---|---|---|---|
| BABILong, average of qa1 to qa5 at 10M, official scorer | 97.2 | 76.6 for the best published entry, a model fine-tuned on these tasks | n=100 per task per seed, 3 seeds, deterministic. Scoped at 10M, never "every length". Task-adapted pipeline; our audit of what that contributes is published beside it. | 2026-08-25 |
| BABILong qa3, multi-hop, 2M to 100M | 81.0 / 80.0 / 80.0 / 80.0 / 79.7 | Beyond every frontier window no reader without the store is measurable | Official scorer, 3 seeds, n=300 per scale, about 685 reader tokens per question at every scale. | 2026-08-27 |
| BABILong qa3 at 1M, matched reader pair | 66.7 with the store | 34.4 for the same frozen Sonnet 4.6 alone, +32 points | n=90 item-paired, McNemar p=1.5e-5. Both arms the same model; only the architecture varies. | 2026-07-30 |
| Repo-Evolution: currency in an evolving codebase | 88 frontier reader / 82 8B reader | Mem0 51, git-grep loop 29, notes file 25, flat retrieval 18 | Same model, same questions; gold derived mechanically from git, no judge in the scoring path. | 2026-08-27 |
| RULER variable tracking, 1M to 10M stored | 24/24 at 1M; 16/16 at 2M, 5M and 10M | No frontier comparison: the June cells are retired pending a clean re-run | Exact match, 2 seeds, corpora sha-pinned, about 260 tokens read per query. Scoped to the chain-walk. | 2026-08-24 |
| Supersession: when facts change | 98% current-answer accuracy on 108 update probes | +70.4 points pooled over similarity retrieval | 3 seeds, discordant-pair McNemar, date-obscured chains. A hand-ordering steelman recovers parity, isolating write order as the discarded signal. | 2026-08-24 |
| MuSiQue, beyond the window | 43 to 28 exact match | 19 to 0 for a truncated full-context reader | n=30, 3 seeds, official exact match, no judge. A gentle decline, not a plateau; 5-hop is unsolved. | 2026-08-24 |
| HippoRAG-2 at 1M, identical items | 84.0 | 15.5 for HippoRAG-2, the strongest published structured-memory system | The identical 100 qa3 items at 1M; its token footprint is more than 2,500x ours. | 2026-08-24 |
| Modern retrieval, built by us, on the official 10M cells | 11.3 to 12.3% for the steelman | Against 86 on the same official cells with the store | Hybrid keyword, dense and rank fusion at a 2 to 4K budget, pre-registered with its kill condition, reading 4.9x our tokens. At our budget it scores 3%. | 2026-08-27 |
| Real novels beyond the reader budget | 38.6% pair accuracy | 20.1% for positional truncation | Effective n=63 book-pairs, same reader and 24K budget in both arms. Scope is retrieval, not multi-hop reasoning. | 2026-09-02 |
| Consolidation, write-time aggregation gate | +22.8pp (95% CI +8.1 to +38.2) | A compute-matched language-model counter emitting the identical sentence | Covered aggregation questions. The same contrast runs -3.3pp on trend and is neutral on single-episode controls. Single reader, single scale, synthetic instrument. | 2026-08-14 |
| Weight path, retention (lab) | 47 to 75 points | More prior knowledge retained than the no-adapter control, at matched-or-better acquisition | Not shipped. What ships is the store, retrieval and consolidation over a frozen model. | 2026-08-24 |
| Published null: within a window, single shot, static corpus | +0.00pp, p=1.0 | A frozen frontier model with strong retrieval ties us | Pre-registered, losing condition published before the run. The one regime where structure has nothing to add. | 2026-08-24 |
| Discovery channel: how long the answer sat there | 29 of 155 realized imports, median lead time 12.5 years | A median 72% of the delay from problem to solution was reducible in principle (n=48) | Those 29 sat in a skimmable top-40 slate more than a year before the field imported them; abstract-level ranks, so a floor. The reducible share is a connection left unmade. | 2026-08-09 |
At capture, work is distilled into knowledge objects: claims with their reasoning, evidence and provenance. Encoding is gated by prediction error: what contradicts or extends the store writes strongly; the redundant fades. When a new claim lands on an entity that already carries one, the two are linked at write time and the old claim is marked superseded. At read time, retrieval walks the chain to its head, so a superseded value cannot outrank its replacement however similar it is to the question. Offline, consolidation compacts specifics into structure: fast capture, slow integration, the division complementary learning systems theory describes.
The read path is where the cost story lives. A query goes to the router, retrieval walks each supersession chain to its head so a stale value never reaches the model, and what arrives is a few hundred tokens of resolved evidence rather than the history. That payload stays near the same size whether the store holds a megabyte or a career, which is why accuracy and cost stay flat as stored history grows. The flatness is scale-invariance by construction, not a claim that reasoning survives a hundred million raw tokens unaided.
Six BABILong figures appear across our pages. All six are real and scoped; this is the scope of each, and the one that is the headline.
| Number | What it measures, and its scope | Ledger | Role |
|---|---|---|---|
| 97.2 | Average of qa1 to qa5 at 10M, official scorer (97.5 is the 3-seed judge figure). At 10M only; task-adapted pipeline. | row 2 | the headline |
| 24.7 | Zero-adaptation audit: the shipped generic pipeline, no task adaptation, average at 10M (qa1 4.7 / qa2 4.3 / qa3 6.3 / qa4 45.7 / qa5 62.7); pooled 3 seeds, n=300 per task. | row 3 | published audit |
| 80.0 | qa3 (multi-hop) flat from 2M to 100M, official scorer, 3 seeds, n=300 per scale. | row 4 | the scale claim |
| 87.0 | qa3 on the official HuggingFace cells: 86 / 87 / 87 / 84 / 86 at 128K / 256K / 512K / 1M / 10M; n=100 per cell, seed 42, single seed. | row 1 | ledger only |
| 66.7 vs 34.4 | Matched-reader pair at 1M (qa3): the same frozen Sonnet 4.6 with and without Sapience, n=90 item-paired. | row 14 | the 1M pair |
| 68.9 vs 31.1 | An earlier n=45 half of that same pair (+37.8), superseded by the n=90 pooled figure. | retired list | archive only |
Two qa3 values at 10M exist on purpose: 86 in row 1 (official HuggingFace cells) and 80.0 in row 4 (generator-built corpora); same task, same scorer, different corpora. Every BABILong number here comes from a task-adapted pipeline, in the same sense that the fine-tuned entries on that leaderboard are task-trained; the 24.7 audit measures how much that adaptation contributes.
Public coding benchmarks resolve inside a single context window, so they cannot tell a system that accumulates knowledge about a codebase from one that re-derives everything per run. Repo-Evolution targets that regime directly: an agent experiences the real git history of four production repositories as ordered time-slices, then answers which value is current and which value replaced which, with gold derived mechanically from git.
| Arm | Accuracy | What it is |
|---|---|---|
| With Sapience, frontier reader | 88 | Each change is recorded as a supersession edge at write time, so a currency question becomes a traversal to the head of a chain. |
| With Sapience, 8B open reader | 82 | Swap in a model a hundredth the size and the margin holds: the store resolves the answer, the model transcribes it. |
| Mem0 | 51 | The leading memory tool. |
| Agentic git-grep loop | 29 | Search over the real history, the tooling a good engineer would build. |
| Running notes file | 25 | The pattern most teams actually use. |
| Flat retrieval over the full history | 18 | Write order is discarded at ingestion; similarity search cannot recover it. |
A task-level pilot, on whether the same agent makes a real code change more cheaply with an accumulated memory, has produced two single observations with a split sign and no result yet. The powered, pre-registered run reports either way. Detail on the coding evidence page.
Two regimes matter beyond the leaderboard. First, multi-hop reasoning once the evidence outgrows the window: iterative retrieval over the store declines gently while a truncated reader goes to zero. Second, reach: connections valuable precisely because they lie outside what has ever co-occurred, which every co-occurrence instrument is blind to by construction. Within its own horizon dense retrieval out-recovers our channel, so this is a complement; its value is the stratum retrieval cannot enter at all.
The whole system runs on machines you own, with open models a hundredth of frontier size. Reads scale with the query, not the history: a few hundred tokens of resolved evidence per answer instead of hundreds of thousands, about 1,800x cheaper per query on the licensed figure. The alternative is dumping the store into the context, which collapses as the store grows past the window (78.3 to 41.7 to 13.3 on lookup probes) while retrieval over the persistent store holds 100% from about 961x fewer tokens per answer. Latency behaves the same way: about 1.4 seconds per query against about 10 seconds for the frontier model alone at 1M on a warm cache, roughly 7x, flat as history grows. The store grows with what you keep, not with what you read. Size is a storage question, not a reasoning-cost one.
Matched protocol (same model, judge and items per cell), multi-seed, calibration holdouts excluded from every headline, cross-judge audits, item-paired statistics, mechanical ground truth wherever the task permits, and pre-registration with published kill conditions. Every number traces to a result file and the script that produced it. Reproduction bundles on request.
Investor reading room: room.spnc.ai