Skip to content

✨ Recall is asked from a scene, and says what each memory's scene was - #121

Open
ebigunso wants to merge 4 commits into
feature/2026-09-21/scene-on-writesfrom
feature/2026-09-21/scene-on-retrieval
Open

ebigunso wants to merge 4 commits into
feature/2026-09-21/scene-on-writesfrom
feature/2026-09-21/scene-on-retrieval

Conversation

@ebigunso

Copy link
Copy Markdown
Owner

Task_3 of the scene plan (stacked on #120). It completes the scene slice.

Asking

  • A retrieval gives a Scene (the same value a write records) and an optional topic. Only the time is required and it defaults to now. The application resolves nothing and the facade exposes no lookup (ADR-D-0029, ADR-I-0020).
  • A participant given by key cues its notion; by name, every notion currently known by exactly that name; by description, and the setting's words likewise, whatever memory holds that matches in content. With no topic, no topic is embedded.
  • Roots the scene names enter before roots from content hits, in the scene's own order, each notion once, at full strength on the existing scale; the trace reports any root the budget dropped and marks these roots as named by the scene, never as vector hits. The existing measured selectivity bounds them like any entity root, which is what gives it its production input back.
  • A retrieval that gives only a topic selects the same memories, in the same sections and order, from the same query text as before.

Reporting, in the result itself

  • Every admitted memory reports the scene it was recorded in: an episode its own, an observation its episode's, an interpreted memory the scene of every experience it rests on, admitted or not. A memory with no recorded experience behind it says so explicitly. A source that is missing, or that the character was told to forget, is reported as not available with the reason and no contents.
  • What each reference resolved to is part of the result: resolved, ambiguous with each notion it could mean, unknown, or taken as a content cue with what it reached. Knowing that you do not know someone is knowledge the character acts on; a description is never called unknown, because that needs a measured floor.
  • The present scene is echoed as given. An unset part means not given, and no scene is ever complete: there is no completeness flag and no list of absences.
  • Nothing is withheld because of a scene and no verdict about who may hear what is computed (ADR-D-0038). The setting key and custom values are reported back and cue nothing yet.

Evidence: internal Tier D review approved with no findings, before and after the Tier A round; the Tier A review judged the result fit for a model to be discreet with, after the absence list was deleted and the reference outcomes moved out of the optional trace. fmt, check, clippy with warnings denied and the full suite pass for the worker and independently for the reviewer.

🤖 Generated with Claude Code

ebigunso and others added 4 commits September 21, 2026 06:44
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Remove the absence-derived scene API and preserve the weakest search verdict with its own counters. Document inert setting keys and custom values; verify reference outcomes with tracing disabled.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant