Nemosyne — Notes https://nemosyne.ai/notes/ Writing on institutional memory, citation and permissions. en The seven limits we published before anyone asked https://nemosyne.ai/notes/seven-limits/ https://nemosyne.ai/notes/seven-limits/ Fri, 11 Sep 2026 00:00:00 +0000 Every system has a list like this. The only question is whether the buyer reads it from you, or assembles it from a security review, a failed pilot and a conversation you were not in. What a citation actually proves https://nemosyne.ai/notes/citation-proves-provenance/ https://nemosyne.ai/notes/citation-proves-provenance/ Fri, 11 Sep 2026 00:00:00 +0000 Every sentence carries a citation, and every citation opens a real passage. That is provenance, and it is absolute. It is not the same as the passage supporting the sentence — and the difference is the seventh limit we publish. Why we reconcile disagreeing values without a model https://nemosyne.ai/notes/reconcile-without-a-model/ https://nemosyne.ai/notes/reconcile-without-a-model/ Fri, 11 Sep 2026 00:00:00 +0000 Two documents say different things about the same number. Asking a model which is right is the expensive answer, the unrepeatable answer, and usually the wrong one. Three deterministic rules settle almost all of it. Entity resolution across four systems that never spoke https://nemosyne.ai/notes/the-join-across-four-systems/ https://nemosyne.ai/notes/the-join-across-four-systems/ Fri, 11 Sep 2026 00:00:00 +0000 Every system in your firm knows half of it. None of them knows that its half and another system's half are about the same company, and no amount of search will tell them. Permissions that can only narrow: the intersection rule https://nemosyne.ai/notes/permissions-can-only-narrow/ https://nemosyne.ai/notes/permissions-can-only-narrow/ Fri, 11 Sep 2026 00:00:00 +0000 Visibility is stamped at capture and composed by intersection at every derivation. There is no step in the pipeline where a permission can widen — which is a stronger statement than a policy saying it should not. Search finds what was written. Memory notices what a senior colleague would notice. https://nemosyne.ai/notes/not-search/ https://nemosyne.ai/notes/not-search/ Fri, 11 Sep 2026 00:00:00 +0000 The difference is not relevance ranking. It is that one system holds your work as text and the other holds it as structure — and some of the most valuable findings are about things that are not in any document at all. What “read once, write down what it learned” costs, and why we pay it at write time https://nemosyne.ai/notes/write-time-cost/ https://nemosyne.ai/notes/write-time-cost/ Fri, 11 Sep 2026 00:00:00 +0000 Compaction here is literal. The expensive thinking happens on arrival, which makes reading cheap, makes the pricing honest, and makes the first few hours slower than anyone demoing would prefer. An ontology is a vocabulary, not a schema: how packs work https://nemosyne.ai/notes/ontology-not-schema/ https://nemosyne.ai/notes/ontology-not-schema/ Fri, 11 Sep 2026 00:00:00 +0000 The difference is commercial, not academic. A schema is something you migrate to. A vocabulary is something you extend — and the extension belongs to you. Customer concentration, three documents, four months https://nemosyne.ai/notes/concentration-three-documents/ https://nemosyne.ai/notes/concentration-three-documents/ Fri, 11 Sep 2026 00:00:00 +0000 No single document is alarming. The pattern across three of them is the finding — and it exists only if something is holding all three at once. Every firm and figure below is invented. Prompt injection: a structural defence, and no detector https://nemosyne.ai/notes/prompt-injection-structural/ https://nemosyne.ai/notes/prompt-injection-structural/ Fri, 11 Sep 2026 00:00:00 +0000 A classifier that recognises injection attempts is a classifier an attacker can test against until it does not. We would rather bound what a hostile document can cause than play that game and claim we win it.