Shomei admits verbatim, byte-grounded spans instead of asking a model to smooth a conversation into new prose. The source stays attached; unsupported detail stays out.
Memory that can carry weight
The moment an agent acts on what it remembers (a dose, a contract term, a balance), “usually right” is disqualifying. And an invented memory is worse than an invented answer: the answer disappears after the conversation; the memory persists, resurfaces later as established history, and shapes the next recommendation, approval, or decision.
In our benchmark the same model proposed memories from the same conversations. Ungated LLM extraction fabricated roughly 1 in 50 remembered facts; Shomei’s deterministic admission gate allowed 0 through across 76 tested cases. Every stored memory must point back to what was actually said, or the system abstains instead of quietly turning a guess into history.
~2%
≈ 1 in 50 stored facts fabricated (95% CI 1.5–2.9%), and it compounds
Caveat: n=500 conversations, preregistered; fabrication was scored by a single neutral gpt-4o judge against the transcript, not human-adjudicated ground truth.
0
fabricated content admitted (0 of 76)
Caveat: same model, corpus, preregistration, and single gpt-4o judge; zero describes admitted fabricated content in this committed run.
Shomei does not trust the model’s claim. It requires the source.
A per-fact rate looks different once memory accumulates. At the ~2% rate we measured for ungated extraction, the odds a single customer’s memory holds at least one fabricated fact cross more-likely-than-not by the ~50th fact and reach ~88% by the 100th (projected from the per-fact rate). Shomei’s gate held at 0 of 76 tested: it only stores what was actually said.
Speed · query at 100k
~5.7× faster
The optimized scan returned byte-identical output, with a SHA-256 match on all 26 queries.
Caveat: p95 over 26 near-max samples in a single run; elapsed milliseconds are approximate, while the speedup and byte identity are the robust claims.
Deletion · 100k tier
0 readable residue
Erase a subject and a forensic byte-scan finds nothing readable; a co-tenant’s 40,000 memories remain untouched.
Caveat: 100k = 10 tenants × 10k, not 100k per tenant; 2 crash-under-load scenarios were skipped.
Where we win · long-history recall
69.2 vs 54.9
On LongMemEval, governed long-history retrieval beats brute-force full-context by +14.3 pts answerable, on a neutral GPT-4o judge. Retrieving the right span beats stuffing everything into the prompt.
Caveat: n=91 answerable (a sample, not the full 500); a neutral gpt-5.4-mini judge corroborates at 65.9 vs 54.9; the first reranker pass measured cosine (onnxruntime absent).
Where we lose · LoCoMo
0.38 vs 0.54
On LoCoMo multi-hop, governed recall currently trails plain dense RAG. We publish the loss because a benchmark page that only wins isn’t one.
Caveat: recall uses an approximate token-containment matcher (threshold 0.55); ~84% of the gap is extraction coverage, not ranking.
Fabrication: rates are single-LLM-judge (gpt-4o) estimates, not human-adjudicated ground truth.
Recall: LoCoMo recall uses an approximate token-containment matcher. Every headline above links directly to its committed artifact so the result can be inspected and reproduced.