Watch a fact try to get into memory.

Most agent memory stores whatever the extractor emits, then hopes retrieval catches the lie. VeriMem puts a gate at the write boundary. Pick something to feed the engine and follow it, checkpoint by checkpoint, until it is admitted — or turned away. Flip X-RAY for the real thresholds, files and flags.

Every reading on this bench is a real, self-run number from the open repository — the entailment score scale (0–100, write threshold 40), the trust collapse, the abstention floor, the NET(λ) counts. Nothing here is third-party audited, and the page says so wherever a figure appears.

▸ Feed the engine — choose one

λ cost of a wrong answer 2×

00 standby

Pick a scenario, then press Run.

The bench above is the engine's real path — extraction, admission, recall. Each chamber is a module in the open repo; each reading is a number you can reproduce. X-RAY shows files, thresholds, flags.

idle · no fact in flight

NET(λ) at scale · HaluEval QA · 300 probes

VeriMem · floor τ=0.8 +0.58

mem0 · as shipped 0.00

At λ=2: as-shipped mem0 breaks even and goes net-negative beyond — 100 fabricated answers on the unanswerable half. VeriMem stays positive to λ≈45.

What runs, and what is research

A trust product that hides its own switches would be self-refuting. Here is exactly which stage is on by default, which you opt into per call, and which ships default-OFF pending the external scrutiny this is meant to invite.

Extraction Opt-in tooling
Turns raw material into candidate facts: extract_text (pdf/docx/html/epub, zip-safe) and ingest_conversation (LLM proposes → gap-fills → consolidates, provenance conv:<id>). Extraction is generative — it can dream. That is exactly why the gate sits after it.
L1 lexical screen Core · always on
~13 ms, no LLM. Downgrades unsupported "it works / verified / done" self-claims that carry no evidentiary anchor. engram/anti_confab_gate.py
L4 entailment gate Opt-in · the moat
Admits a write only if its cited source actually entails it — source⊢fact on a 0–100 scale, write threshold 40 (answer-path 85), discriminator AUROC 0.971 on SNLI and 0.992 on realistic confabulations, judge-independent. engram/grounding_gate.py
Source-trust Research · default-OFF
Two channels (agreement + use-outcome), trust = their minimum, prior 0.5, quarantine below 0.25; independence collapses colluders to one witness; audit-deconfounding (P88) demolishes a cartel 0.90 → 0.20 on a real corpus, 3/3 seeds. engram/source_trust.py
Epistemic + composition + self-provenance Research · default-OFF
Derived facts pass the same gate, get proven / unbeaten(bound) / refuted labels (refuted is absorbing), cite their parents (derives_from, retractable), and are signed actor:* so the engine can never testify for itself (alarm past 50% self-writes).
Abstention floor τ Core · the point
At recall, if the best evidence is below τ the engine answers "I don't know" instead of returning its nearest neighbour with false confidence. Decision rule: answer iff p > λ/(1+λ). Calibrated to operate at the declared risk (TCE ≤ 0.011 at λ ∈ {0.5–9}).