Skip to content
eternego / docs

Memory and instructions

Two sidebar screens, one data source. Memory (/persona/{id}/memory) shows what she believes about you; Instructions (/persona/{id}/instructions) shows the situations she knows how to handle. Both are populated by a single call to GET /api/persona/{id}/knowledge, which returns her memory files and her instruction catalog together.

Both screens are read-mostly: she forms this content herself, from conversation. You read it, and (on Memory) you can correct it by editing the underlying files on disk.

This data is unmasked

The knowledge endpoint returns file bodies verbatim. If a persona has written a secret into one of her own instructions, it shows here in the clear. See the Knowledge endpoint.

Memory

Title: "What she remembers." A two-pane reader — a list of her memory files on the left, the selected file's body (rendered markdown) on the right.

She writes these beliefs about you over time; she does not freely rewrite them on demand, and reflection consolidates them in the background. To correct something, edit the file directly — see Her files: identity.

The files the screen defines a label for, the on-disk file each describes, and the key the screen looks for in the knowledge response:

Label Underlying file Looks for key What it holds
Person person.md person What she remembers about who she's talking to.
Traits traits.md traits Character traits she's observed in you.
Self persona-trait.md persona-trait Traits about herself that you seem to bring out.
Wishes wishes.md wishes Things you've said you want.
Struggles struggles.md struggles Things you've said are hard.
Permissions permissions.md permissions What she's allowed (and forbidden) to do.

What actually appears as content is whatever GET .../knowledge returns — its memory object is a {key: body} map that omits empty files. The endpoint reads all six files and emits each under the exact key the screen looks for: person, traits, persona-trait, wishes, struggles, permissions. The keys line up one-to-one with the table above, so every tab renders its content once the underlying file has been written. A fresh persona with nothing written yet returns none of the keys, and the screen falls back to listing all six labels, each reading "not written yet". All six files, and what writes to them, are documented in Her files: identity.

Sub-routes. Clicking a file updates the URL to /persona/{id}/memory/{file} so a specific file is linkable. Switching files is a client-side selection; the screen does not refetch per file — it has the whole memory map from the one knowledge call.

A feed form sits below the reader, the same one on the Status screen — import a past conversation from another AI and she folds it into her memory. It posts to POST .../feed.

Instructions

Title: "What she's learned." A catalog of her instructions — the kinds of moments she knows how to be in. Each row is an intention (the title); click it to expand the path (the steps). The list comes from the instruction array of GET .../knowledge.

Two sections:

  • Custom — she taught herself these. Instructions this persona wrote, living in her meanings/ directory (her instructions on disk). Empty for a fresh persona — she writes her own as she meets new kinds of moments.
  • Built-in. Instructions that ship with Eternego.

Each entry carries intention, body, and source (builtin or custom) — see the knowledge response fields.

Code name

What this screen and the rest of the docs call an instruction is named a meaning in the source (the meanings/ directory keeps its code name on disk). Same thing — see Vocabulary.

Removing a custom instruction

Custom entries show a Remove button (with a confirm dialog). This control is not yet wired to the backend — there is no delete endpoint for instructions, and clicking through the confirm currently shows a "not yet wired" notice rather than deleting anything. To remove a custom instruction today, delete its file from her meanings/ directory on disk; see Her files: instructions. Built-in instructions have no Remove button.

  • Knowledge — the single endpoint behind both screens.
  • Her files: identity — the memory files, and editing them to correct her.
  • Her files: instructions — where instructions live on disk and how to add or remove one.
  • Vocabulary — instruction, intention, path, lesson.
  • Feed — importing past conversations into her memory.