Extension Output Showcase

Three ways to use this page

• Show the page directly. Open it full-screen and scroll as you talk — every block below looks like the app's real output.
• Reproduce it live. Each section ends with a Try it line giving the exact commands, so you can recreate the demo on camera in a fresh chat.
• Make a demo character first. Create a throwaway character and start a new chat before recording: an empty memory store and an empty log folder mean nothing personal can surface while you demonstrate.

1.Command reference at a glance

Everything below is typed into the normal chat box. Nothing starting with a slash reaches the model — it returns instantly and can't be hallucinated.

2.Memory: saving, keyword recall, meaning-based recall

2.1 Saving a memory — explicit keywords

2.2 Saving with automatic keywords

No keyword list given, so the extension picks them — words wrapped in *asterisks* take priority.

2.3 The memory store

(Saved memories:
[1] My sister is training for a charity hike in the Peak District in May.
[2] I'm rebuilding a 1978 cassette deck in the workshop — belts and pinch roller done.
[3] Prefers green tea over coffee, never coffee after 4pm.
[4] Works shifts, so evening chats fit best.)

2.4 Keyword recall — what the model actually receives

Weeks later, the trigger word appears and the memory is injected into that turn. This is the block the model sees above the message:

[Long-term memory for this character:]
- My sister is training for a charity hike in the Peak District in May.

2.5 Meaning-based recall — no keyword needed

Same memory store, different wording. Nothing here matches a keyword literally, but the local embedding model recognises the subject:

[Long-term memory for this character:]
- My sister is training for a charity hike in the Peak District in May.

3.Continuity between sessions

Open a new chat and the tail of the previous one is carried in, along with a nudge not to open with the same stock greeting:

[Previous session context: last chat was 3 days ago]
Sam -> Aria: how's the hike prep going?
Aria -> Sam: Still on the training plan, last I heard…
Sam -> Aria: I'll sort the sponsorship page at the weekend.
Aria -> Sam: Good — do the map first, it does more work than the paragraph.
Sam -> Aria: night then.
Aria -> Sam: Night, Sam.

[Open this session naturally - vary your greeting. Your mood is good; relationship with the user is 72/100. Don't repeat a stock opener.]

Why it matters on camera: the first line references the previous session rather than saying "Hello! How can I help you today?" — that's the whole point of the carry-over.

4.Persistent state: mood, style, relationship, profile

All four survive restarts and are injected as context, so they behave like state rather than prompt decoration.

/mood
(Mood: good (score +6), energy medium, patience high)

/style
(Writing style:
Tone: casual | Swearing: mild | Sentences: medium | Humour: dry | Catchphrases: none)

/relation
(Relationship with Sam: 72/100)

/profile
(User: Sam.
User profile:
[1] Works shifts, so evening chats fit best.
[2] Prefers the plain answer first, context after.)

What the model sees each turn:

[Mood state for this character:]
- Mood: good (score +6)
- Energy: medium
- Patience: high
- Relationship with user: 72/100
- Recent tone: good

[Writing style for this character (keep this voice):]
- Tone: casual
- Swearing: mild
- Sentence length: medium
- Humour: dry

[Facts about Sam (the user, shared across all characters):]
- Works shifts, so evening chats fit best.
- Prefers the plain answer first, context after.

The state card (panel readout)

Character: Aria
Mood: good (score +6) | Energy: medium | Patience: high | Relationship: 72/100
Style: Tone: casual | Swearing: mild | Sentences: medium | Humour: dry | Catchphrases: none
Active user: Sam
User facts: Works shifts, so evening chats fit best.; Prefers the plain answer first.
Memories stored: 4
Rolling summary: - Sam is rebuilding a 1978 cassette deck: belts and pinch roller done…

5.Rolling summary (and the rejected one)

Long chats get a background summary so the assistant stays coherent beyond the context window. It is generated off the live reply path and checked before it's kept.

What /summary shows when the summary is good

(Rolling summary:
- Sam is rebuilding a 1978 cassette deck: belts and pinch roller replaced, head alignment left for the weekend.
- His sister is training for a charity hike in the Peak District in May; he's running the sponsorship page.
- Long-running thread: sci-fi novels with unreliable narrators - I still owe him a recommendation.
- He prefers the short answer first, then the reasoning.)

What happens when a generation comes back bad

User: CURRENT_LOCAL_DATETIME=2026-09-24_21-04-11
 Sam: evening ( Sam: evening ( Sam:
 evening ( Sam: evening ( Sam: evening
 ( Sam: evening ( Sam: evening ( Sam: …

Same bullets as above. The bad generation is
discarded, the stored summary is untouched,
and the next update tries again.

Demo line: "if the model has a bad moment, the fix is on the extension's side — a repetition loop never becomes the assistant's memory of the chat."

6.Summariser output

A bare command in the chat box turns into a summary with a fixed shape: one summary section, five keywords, written in the assistant's own first person.

Before and after the output contract

Aria: Right, let's see what the archive says…
[Summary]
The conversation has centred on a workshop
project, where Sam and Aria explored the
repairs. They discussed how Aria thinks the
belt replacement went well and Sam shared
his views on the deck.
[Keywords]
- workshop
- belt
- repair
- project
- deck
- cassette
- weekend
- alignment
- tools
- patience

[Summary]
Sam came in with the cassette deck finally
running — belts and pinch roller done…
[Keywords]
- cassette-deck
- charity-hike
- peak-district
- sponsorship-page
- unreliable-narrators

7.Session diary

/diary writes a dated entry to that character's folder; /diary read shows the file. Built from the session itself, so it always has content:

## Diary - Thu 24 Sep 2026, 21:12
**Aria** · 14 exchanges with Sam · 19:58 → 21:12 (~1h)

- Mood: good (+6) · energy medium · patience high
- Relationship with Sam: 72/100 (close)
- Mood through the session: good throughout (+2 → +6)

- Threads: cassette, workshop, weekend, alignment, sister, charity
- Line that stayed with me: "I'd rather leave the alignment until Saturday than rush it and snap a head I can't replace."
- Saved to memory:
  - I'm rebuilding a 1978 cassette deck in the workshop — belts and pinch roller done.
  - My sister is training for a charity hike in the Peak District in May.

- Reflection: - Sam is rebuilding a 1978 cassette deck: belts and pinch roller replaced, head alignment left for the weekend.

Worth saying on camera: slash-command turns are skipped, so a command and its reply never end up quoted as the "line that stayed with me" or the last thing said.

8.Verbatim log recall

Ask what was actually said earlier and you get the literal lines with real timestamps — no paraphrasing, no guessing.

Why it's demo-friendly: the recall replaces the reply outright, so the model can't pad it, paraphrase it or invent a line that was never said. Previous recall dumps are filtered out too, so recalling twice doesn't nest one dump inside another.

9.Time injection & reminders

Appended to every message — time of day, in words

evening Sam, how did the weekend go?

[Time context - injected, not message text. It is evening on Thursday. Speak about time in broad terms like that; give an exact clock time or date only if asked. Never quote or repeat this line.]

No clock and no calendar date sit in an ordinary prompt, so there is nothing for a model to copy — the assistant just knows it's evening, the way a person would.

When you actually ask

… Never quote or repeat this line.
CURRENT_LOCAL_TIME=22:57 CURRENT_LOCAL_DATETIME=2026-09-24_22-57-30 UTC_OFFSET_MINUTES=60 CURRENT_UTC_DATETIME=2026-09-24_21-57-30]

"what's the date?", "what day is it?", "what did we say earlier?", "was that last night?" add just the calendar date; a direct time question adds the clock. So accurate answers come from real values, while the default state stays quiet.

Shape, order and tiering are all deliberate. The block used to lead every message with a full timestamp, which read as the opening of the user's own text — models copied it back as a dateline at the top of their replies, and led with UTC so "the time" came out an hour behind the local clock. It now trails the message, states the time of day in words, and only carries exact values when they were asked for. Field names are unchanged, so the persona rules that refer to them still apply.

A delivered reminder

Reminders are stored in a plain .ini (weekly / monthly / yearly / one-off), delivered as their own assistant-side message, and throttled to once every 20 prompts by default so they never spam. There's an on/off switch in the panel.

10.File workspace: commands and the panel

Real file operations in a shared folder, with no model involved. Ideal for a video: instant, verifiable, and impossible to hallucinate.

Line-targeted editing and undo

• Every modifying command snapshots the previous version first, so /file undo steps back — call it again to step back further.
• Partial names resolve themselves: notes finds notes.txt, searched recursively through subfolders.
• Paths are locked to the workspace folder, so nothing can be written outside it.
• Several commands can be pasted at once, one per line.

Doing it by hand — the File Workspace panel

The commands are for the model. For you there's a panel (its own extension, its own section in the Extensions tab) that drives the same engine — what you click and what the AI types do exactly the same thing to the same folder.

What attaching looks like

Your message carries the file's contents, so it's part of the conversation and part of the transcript — no hidden prompt injection and nothing to wonder about.

11.Chat export

/export
(Exported 42 exchanges to: …/long_term_memory/_exports/Aria-20260924-2112.md)

/export html
(Exported 42 exchanges to: …/long_term_memory/_exports/Aria-20260924-2115.html)

A clean transcript with the injected context blocks stripped out — no model involved. Show the file opening in a browser if you want a "look what came out" moment.

12.Reply clean-up, before & after

Models like to imitate whatever they can see in the prompt. These are the real shapes of that problem, and what the reply looks like after the extension's clean-up pass. Good "the unglamorous work" segment.

Copying the injected timestamp — a dateline

2026-09-24. You made it through the night. I'm here whenever you want to pick up where we left off.

You made it through the night. I'm here whenever you want to pick up where we left off.

Copying the injected timestamp — woven into the sentence

Cutting this one would leave a broken fragment, and the raw value a model reaches for is the UTC one, so the stamp is rewritten as the local clock instead:

21-27-38 is a bit late for a "hi" but I'm glad you're here. (local time was 22:27)

22:27 is a bit late for a "hi" but I'm glad you're here.

A bare HH:MM is left untouched, because that is the shape a real time answer uses — the repair only applies to the macro's own value shape (21-27-38, 2026-09-24_21-27-38, 22:27:38).

Signing the reply with its own name or version

Aria: That's a fair point, but the map still does more work than the paragraph.
1.0. That's a fair point, but the map still does more work than the paragraph.

That's a fair point, but the map still does more work than the paragraph.

Parroting an injected context block back as dialogue

[Mood state for this character:]
- Mood: good (score +6)
- Energy: medium
- Patience: high
- Relationship with user: 72/100

Evening, Sam.

Evening, Sam.

Keeping the paragraphs the model wrote

The deck's running.

Head alignment next, then it's done.

The deck's running.

Head alignment next, then it's done.

Also handled: stray HTML tags the model echoes, a reply reduced to a bare status token, and time macros read out loud by the speech engine.

13.Voice: speech and the voice loop

Both are fully local — a small neural voice model and a portable Node runtime, no PyTorch and no cloud service.

Text-to-speech panel

persona_voices.json

{
  "Aria": "bf_emma",
  "Cole": "bm_george"
}

Before speaking, the text is cleaned: thinking blocks, leaked time macros and HTML entities are stripped, so the voice never reads ' or a timestamp out loud.

Voice loop (early stage)

One panel starts a hands-free loop: microphone → speech-to-text → assistant reply → local speech → speaker. It's start/stop only for now — frame it as "the loop works, the polish comes next".

Camera tip: audio can't be shown on a slide — record this live. Hit play on two different characters to prove the per-character voice, that's the part that reads instantly on video.

14.Image generation

Local Stable Diffusion on Vulkan, with the checkpoints bundled. No PyTorch stack, no separate SD web UI, no API.

The AI can draw on its own

Say in the prompt that it may draw, then the model emits an image tag inside its reply and the tag is swapped for the finished picture:

• Panel controls: model, prompt, negative prompt, size, steps, CFG, seed.
• Each model remembers its ideal settings, applied automatically when you pick it.
• Drop another checkpoint into the models folder and it appears in the dropdown — no code change.
• VAE tiling on by default, so 1024×1024 fits on a 16 GB card.

15.The Extensions tab

All of the above lives in one top-level tab, with each panel collapsed until you open it.

The two remaining extensions are hook-only — the summariser and log recall add commands and behaviour rather than controls, so they have no panel. The four core panels are always loaded and can't be switched off by accident from the extension list.

16.Reproduce-it checklist

Paste-ready. Start a fresh chat (ideally with a throwaway demo character) and work down the list.

Formats on this page come from the extensions' source, not from memory: command replies, injected blocks, the diary layout, the recall line format and the panel names were all read back out of the code. The content is invented for demonstration. Names used: Aria (assistant) and Sam (user) — swap them in your head for whichever character you're showing.