Eve library

Chapter 8 · Context, memory and evidence · brief · 10 September 2026

Guide

Revision 1

On this page

The second chapter of the How Eve works track, rewritten in the Eve edition on 10 September 2026. It follows the plan agreed on 9 September and recorded in EVE-V1-REDESIGN-TODOS.md. The receipt for the delivery is under verification/; this file records what the chapter is for and why it is shaped this way.

The job#

Chapter seven ended with an authorised turn carrying an object. This chapter is about everything else that reaches that turn: the selection on the page, the conversation so far, what may be remembered across turns, and what actually supports a factual claim. They look interchangeable in a prompt, and they are not.

Running example. An operator resumes an earlier thread in another workspace, with a draft, a documentation excerpt and a receipt.

Three learning promises (the cover)#

  1. Four layers, four owners — page context, conversation, memory and grounding, each with its own purpose and expiry.
  2. What may be remembered — a governed read, a governed write, and a forgetting that survives the next import.
  3. What supports a claim — claim-level support, a bounded corpus, and the difference between a document and a fact.

Three things to remember (the close)#

  1. Every layer has an owner. Being assembled into one prompt changes none of its purpose, ownership or expiry.
  2. Recall and deletion are governed. Relevance is not permission, and a removed row is not a deletion without a fence behind it.
  3. Support is per claim. One statement, one source record, one state a reader can check.

Bridge to chapter nine. The boundary that decides whether any of this may be acted on.

Sections#

Four sections, opened by a section-divider that names all four.

Section Slides
Four layers, four owners purpose and expiry, the layer stack, one turn's three roles, the store map
The conversation append and compact, what compaction may not do, resuming elsewhere
Memory, and forgetting governance at the point of use, a governed read, three states
Evidence and retrieval grounding, per-claim support, the corpus, documented-or-done, the dense index, four failures

Treatment decisions#

  • The dense-index slide is the library's first stat panel. It is the one slide in these chapters with real measured numbers to show — 57,008 rows, 1,024 dimensions, and zero opted-in member pages — each citing the source it was read from. The fourth number is deliberately absent: there is no retrieval-quality result, because hybrid retrieval is not the serving path, and the panel's missing-value marker says exactly that.
  • Forgetting is a state machine. Active, paused and forgotten are three states with four transitions, and forgotten is terminal only because a tombstone makes it so. A flow could not show why the third state needs a fence.
  • Two compare panels, for two genuine two-sided distinctions. What compaction may and may not change; and what a documentation corpus can answer against what only live state can.
  • One legacy graph retired. eve-grounding was the third of the diagrams.json overrides; rewriting it as a step journey leaves nineteen.

What was rebound#

  • Twenty-two slides now live in a new content/66-eve-context.json; they came from 02-eve.json, 03-y-current-eve.json, 21-eve-deepdives.json and 52-eve-memory-detail.json, the last of which is now empty and removed.
  • Six assignment fingerprints re-reviewed rather than re-pointed, across two assignment files, each with a dated note.
  • The Eve-course shape test now accepts a rewritten chapter that opens on its own cover and ends on its own close, instead of requiring a worked example in second position.