CleverMemory
Menu

Research, decisions, and release work

Blog

These notes cover the round trip this product is built around: turning documents into evidence that can be inspected, and turning supported evidence back into documents that show their work. Some articles explain established research. Others record a product decision, a small test, or a public-format choice. Each one should make clear which of those it is.

The Foundations series, parts 1 through 10, is the research spine. Each part takes one field that solved part of this problem, gives you its primary sources, and states what it left unsolved. Part 9 covers the seams: where two of those literatures describe the same phenomenon in terms that do not line up. The series hands over the pieces and their sharp edges. How they assemble into a single record is our implementation, which stays unpublished.

Articles after the series move from the research to the product: what it returns, how it behaves when the evidence does not cooperate, and how any of it would be evaluated. The first of those build notes is scheduled for September 14, with the rest weekly after it. Each appears on its date once its pre-publication review clears; when the next date is confirmed, it appears below.

For product and business leadersWhere this changes a decisionStart with the practical patterns, shown on one published case file, then the evaluation rules that any published result will have to follow. For technical leaders and buildersStart at part 1 of FoundationsTen fields, their primary sources, and the boundaries each one leaves behind.

Articles publish on their listed date. Follow the Atom feed for new notes and release updates.

The next article arrives.

Foundations

  1. Why similarity is not memory Vector retrieval is valuable for discovery, but trustworthy recall needs explicit treatment of polarity, direction, and missing evidence.
  2. A question sets requirements, not keywords A trustworthy memory response has to satisfy every material part of a question and retain proof for each step.
  3. Certain answers: the math of saying "I don't know" Certain-answer semantics explains when incomplete data supports a claim, when absence is not falsity, and why exact counts require a completeness boundary.
  4. The whole sentence informs its parts Event-centered representations keep participants, time, and provenance attached without pretending every sentence is a single subject-verb-object fact.
  5. "Might" is not "is": storing what you actually said A trustworthy memory must preserve negation, possibility, permission, and attribution instead of flattening every sentence into an asserted fact.
  6. Time is two-dimensional Bitemporal records separate when a claim applied from when it entered the system, preserving corrections without confusing history with current state.
  7. The algebra of citations Provenance algebra shows how source lineage can survive joins, alternatives, deduplication, and carefully defined computed answers.
  8. The same number, three meanings Document layout is not presentation to be stripped before indexing. A figure's row, column, and heading are often the only thing separating two identical strings.
  9. The seams between the fields Eight research areas each solved part of this problem in isolation. Where two of them describe the same phenomenon in incompatible terms is where the engineering actually lives.