Memory and Minerva

Mercury's memory features are experience cards, a project notepad and Minerva's room over your saved prompts. Minerva refines a saved prompt only when asked; a refinement lands beside your wording, never over it, and also in the MINERVA tab, a durable per-project feed of refined prompts.

Memory surfaces

/note <text>: Captures a note into the project's journal and rebuilds its notepad.md view, with no model turn.
/tabula: Opens Minerva's room. Its one job is refining the saved prompts, the third tab of the prompts panel opened by /workbench.
/minerva <message>: Turns one message into note operations on the journal with one billed curator call: add, done, pri and refine, never delete.
/submodels: Picks the model and effort for the two sub-model containers, minerva and console.
MINERVA tab: Keeps the project's feed of refined prompts.

Record a project note

To record a project note, follow these steps:

Enter /note <text> with the note's contents.

Mercury appends an event to the project journal and rebuilds the readable notepad without starting a session turn.

Notes live under <config-home>/tabula/, one directory per project, outside the repository, and survive /clear. journal.jsonl is the single source of truth, an append-only event log, and notepad.md is its derived, readable view. Notes carry a priority from the set now, next and later. /note and /minerva are user-private: the line acts on the screen and never enters the session's conversation, and /minerva spends its one call on the Minerva container alone.

Refine a saved prompt

To refine a saved prompt, follow these steps:

Enter /tabula.
Select a saved prompt with ↑ or ↓.
Press Enter.

Minerva refines that prompt, and the result lands beside the original and in the MINERVA tab. Composing is free, and pressing Enter is the one billed call.

The room opens with focus on the saved-prompts list. One ↓ past the list, or Tab, reaches the message box for a direct request. Which prompt a request is about comes from the request itself, a number, an ordinal or a prompt id, and an ambiguous request gets the room's own question asking which prompt is meant. With no Minerva model pinned, the room says so in one line and makes no model call.

Stage a refinement in the main composer

To stage a refinement in the main composer, follow these steps:

Select a refined prompt in Minerva's room or the MINERVA tab.
Press s.

The refined text enters the main composer, the room closes onto it, and nothing is sent until you press Enter.

m on a saved prompt stages its text into the room's message box as an editable draft. Esc from the list closes the room while a running Minerva turn keeps working; two presses of Esc at the working composer interrupt that turn.

Model and effort selection

The two sub-model containers, minerva and console, each take their own model by one precedence: environment pin, then saved pick, then unset.

Environment pin: MERCURY_MINERVA_MODEL and MERCURY_CONSOLE_MODEL take precedence over a saved pick.
Saved pick: /submodels saves the pick across sessions and offers the full catalogue the main /model picker offers.
Unset: No default is derived. A message to an unpinned container spends no model call, and the reply is the line use /submodels to pin one of the available model catalogues.
Effort: e on a model row opens the effort strip with the levels that model offers under the container's call context. Enter saves the level for that container, and Esc keeps the prior one.

Minerva calls with thinking off, so a model whose effort dial is its reasoning dial answers with a one-line receipt and opens no strip. A saved level the picked model lacks runs the model default and says so.

Feature switches

MERCURY_TABULA: On by default. 0 is the only off switch, and off means the store, the three commands, the board and the rail card are absent.
MERCURY_TABULA_MINERVA: Off by default. Arming it is the billing consent for the boot-time curator pass, also a Boot Menu row.
MERCURY_TABULA_DIR: Overrides the store root.

The boot curator pass runs at most once per boot, only when it is opted in, notes exist, the journal advanced since the last run, and the session is interactive; it never runs in a -p run or a daemon worker. Its plans are validated before they apply, and a refinement whose base text no longer matches is skipped, so your edits win.

Sessions, Workshop, Models and effort, Workspace trust and permissions