An AI assistant that remembers everything can still forget what matters. It can fill an answer with recent machine activity while missing the decision that explains why the work exists. It can also find the right document for the wrong person. For a company evaluating AI, relevance and permission are two separate acceptance tests. Passing one does not compensate for failing the other.
My September 4 work on Valeska addresses both problems. This is a research journal from my own working system, not an announcement that a customer's knowledge estate is solved. I want a quieter memory and permissions grounded in the documents' source. I also want the exceptions visible before anyone trusts either claim.
Less noise is a measured result
The wrong default is to treat every recent record as equally useful context. Automated status activity is evidence that machinery ran, not necessarily evidence of a decision. When it crowds out the reason behind a change, an assistant can sound current while being historically wrong. Storing more of it does not repair that distinction.
The recorded recall test used twenty questions. Machine-generated noise among the top eight results fell from thirteen rows to zero. Questions with a relevant hit in those results rose from fourteen to sixteen. That is a useful improvement on this set, with four questions still missing a relevant hit. It is not evidence that irrelevant memories can never return.
The day's related work gives memories a useful life and handles duplicates and replacements when information is recorded. In business terms, yesterday's instruction should not quietly compete with its replacement forever. But removing repetition can also remove a meaningful difference. I regard retention and replacement as decisions to inspect, not housekeeping that becomes correct merely because it is automated.
The source is evidence, not a guarantee
Access control means deciding who may see or change something. An ACL, or access-control list, is a recorded set of those permissions. If AI search maintains a separate account of who can read a document, that account can drift from the place where the company actually manages the file. A convincing answer does not excuse that drift.
The new filesystem connector derives Valeska permissions from the source files' permissions on Linux-style systems. An optional scanning path applies them during indexing. That moves the source into the permission decision rather than inventing every rule inside the search layer. Think of an archive whose search desk consults the restrictions attached to the collection, instead of assuming every visitor may open every door.
The boundary matters. This work imports source-derived permissions; it does not prove exact reproduction of every inherited folder restriction or every corporate identity rule. It is optional, preserves existing owner access, and adds grants rather than replacing the whole permission picture. If source permissions cannot be read, indexing continues with a warning rather than stopping the scan. I cannot call that an independently verified security boundary or a guarantee of continuous synchronization.
Defaults are part of the experiment
Search now defaults to the passage index, which represents smaller sections of documents, instead of silently using the whole-file route after the operator populated the other index. That corrects a practical mismatch between preparation and use. Support for NVIDIA NIM offers another embedding engine, the component that translates text into numbers for meaning-based search. It is an option, not a new requirement.
The memory integration is also packaged for other installations, with remaining installation-specific assumptions documented. Packaging what I use is progress; proving that another operator can use it correctly is the next obligation. Neither an alternate engine nor a reusable package establishes better answers on a new company's documents by itself.
My next tests are deliberately ordinary: repeat recall with unfamiliar business questions, inspect which useful distinctions survive duplicate handling, and compare expected document access with actual search results for different users. Include a changed permission and an unreadable permission record. Until those cases are demonstrated, I have measured recall progress and a source-permission connection, not a blanket promise. That is enough to justify the next bounded test, not enough to skip it.
Historical basis: VALESKA commits d2fcd0c and f759517 (recall and memory lifecycle), a68061e and 5c71cc5 (filesystem permissions and optional scan integration), 5c10e06 (search default), 78dcfc1 (alternate embedding engine), and 57f6e43 (integration package), September 4, 2026. Recall figures are repository-recorded results on twenty questions; permission claims are bounded by the connector and scan implementation, not customer security validation.