Whitepaper
The architecture, argued in full
The whitepaper states the design and its reasoning: why trust in machine memory must be evidenced rather than asserted, and how every trust claim in Cogeto is backed by an artifact you can inspect.
The argument
The memory layers attached to AI assistants are typically opaque accumulators: they remember, but they cannot show what they remember, why they believe it, whether it is still true, whether two of their own sources disagree, or whether anything was ever really deleted. For organizations operating under the European Union's data protection and artificial intelligence regulations, that opacity is disqualifying.
The paper's answer is an architecture in which every trust claim is backed by an inspectable artifact, and its closing argument is the position this product is built on: verifiability, not capacity, is the property that makes machine memory trustworthy.
What the paper covers
Atomic facts with provenance
Memory as discrete, source-linked facts: a verbatim span, an anchored subject, an explicit status, a validity interval.
Verification before storage
Every candidate is judged against its own cited span by an independent pass before it counts.
The suppression log
Unsupported candidates are resolved automatically and recorded in an inspectable log, not queued for triage.
Corpus-wide reconciliation
Entities resolved across aliases and languages, quantities compared arithmetically, verdicts persisted for stability.
Honest retrieval
Fused search with visibility as a precondition, fan out on ambiguity, and unanswerable questions declared unanswered.
Provable deletion
A compensating transaction protocol ending in a signed, hash-chained receipt, re-verified nightly.
The sovereign gateway
A European provider by default, fully local models where required, sensitive entities pseudonymized before any call.
Measured accuracy
The evaluation methodology behind the published per-release trust scores, including the unflattering ones.
Cite it
Golubic, I. (2026). Cogeto: A verifiable, sovereignty-first memory architecture for large language model assistants (v1). Zenodo. https://doi.org/10.5281/zenodo.21702858
DOI 10.5281/zenodo.21702858 · Published under Creative Commons Attribution 4.0.