Two layers
Crystals hold WHAT you know: structured, sourced, never deleted — only superseded, traceably. A soul layer learns WHO you are: your judgment, taste, and patterns.
MEMORY · LIVE · macOS · iOS · Windows
It doesn't just answer. It remembers.
Every other AI starts from zero. Yours starts from everything — and shows you why it believes what it says.
HOW IT WORKS
Memory you can inspect.
The Mirror
Open the SELF page and face it: how the AI understands you, trait by trait, each traceable to the exact writing it learned from. Disagree? Mute it or delete it.
Night Watch
It distills your work overnight — but only proposes. Every candidate lands in your inbox; nothing enters memory without your yes.
Recall with receipts
Three retrieval layers, fixed order: exact match, knowledge layer, semantics. Certainty is never drowned by similarity. Every hit carries source, locator, and confidence.
Nothing enters unseen. Nothing returns without a receipt.
Three live product behaviors, rebuilt from the interfaces in Monolithos. The data below is a sample; the controls and boundaries are the real ones.
SAMPLE DATA · LIVE INTERACTIONTHE MEMORY CYCLE
From raw material to accountable recall.
Capture
Crystallize a document yourself, or let Night Watch find worthy candidates across eight capture paths.
Approve
Automation stops at a proposal. You decide what becomes memory.
Recall
Exact matches run first, then the knowledge layer, then semantic similarity.
Calibrate
Inspect what the system learned about you. Mute or delete the learned trait.
01 · RECALL RECEIPTS
Ask the memory. Audit the answer.
Run the sample query, then open any hit. The retrieval order is fixed, and every result exposes source, locator, and confidence.
MONOLITHOS / MEMORYSAMPLE DATA · LIVE INTERACTION
- 01Exact matchFOUND
- 02Knowledge layerRELATED
- 03SemanticSUPPORTING
Lead with the approval boundary before automation. Trust is the launch narrative, not a footnote.
- SOURCE
- launch/positioning.md
- LOCATOR
- Decision log · lines 42–47
- CONFIDENCE
- 0.96
A local file is not enough. Users must be able to inspect, correct, and remove what the system believes.
- SOURCE
- strategy/sovereignty.md
- LOCATOR
- Principle 03
- CONFIDENCE
- 0.88
Show the origin and confidence beside every recalled fragment so certainty is visible.
- SOURCE
- notes/website-review.md
- LOCATOR
- Paragraph 18
- CONFIDENCE
- 0.79
02 · SKILL ROUTER
One question. The right shape answers.
Monolithos does not pour every recall into the same list. Ask for a judgment and it composes a sourced digest. Ask across a period and the same memory becomes a timeline. Intent and result shape choose the Skill.
CONVERSATIONAL RECALL / ROUTERTHE SKILL IS ROUTED · NOT MANUALLY FORCED
TRY A NATURAL-LANGUAGE QUERY
Summarize my core judgment on AI autonomy.
SYNTHESIZED DIGESTSynthesized from 3 remembered sources
You did not become less ambitious. You moved the ambition from what AI may do to what it must be able to explain.
In February, speed was the throne: a product had thirty seconds to astonish before it had earned the right to explain itself.
By May, a flawless-looking prototype exposed the cost of theater: the interface said it was reading a project while the model received only filenames. “A beautiful lie is still a bug” became the new bar.
By August, the position was no longer anti-automation. It was sharper: let the system discover aggressively, but make consent the last irreversible step.
RESULT-DRIVEN TIMELINECrystal dates are exact. Raw-file dates use modification time and are marked approximate in the App.
2026-0801
Discovery may be automatic. Agreement never is. The system can bring you a candidate; only you can turn it into memory.
memory/sovereignty-contract.md2026-0701
Trust stopped being a promise and became three visible fields beside every hit: source, locator, confidence.
design/recall-receipts.md2026-0501
The project screen looked complete, but the model had received filenames instead of content. From that day, visual success without execution truth counted as failure.
decisions/project-mode-postmortem.md2026-0201
The first instinct was pure velocity: earn attention in thirty seconds, explain the machinery later.
journal/2026-02-14.md03 · THE MIRROR / SELF
See the person your AI thinks you are.
Each trait is learned from your own writing, never from AI output. Open one to see its evidence, then exercise the same sovereign controls: mute or delete.
CALIBRATION / SELF3 TRAITS INFLUENCING DIALOGUE
LEARNED FROM YOUR WRITING
CONFIDENCE0.94
SIGNAL“Cut any feature the user cannot reproduce in the shipped app.”
- SOURCE
- decisions/product-truth.md
CONFIDENCE0.87
SIGNAL“Gold is a signal, never wallpaper.”
- SOURCE
- design/visual-language.md
CONFIDENCE0.83
SIGNAL“Write the red line first; implementation follows it.”
- SOURCE
- operations/review-protocol.md
04 · NIGHT WATCH
It discovers. You decide.
Night Watch can distill writing and discover connections in the background. Its output is always a proposal: rejecting writes nothing; accepting creates the memory.
PROPOSAL INBOXPENDING YOUR REVIEW
01
FOUND IN · journal/2026-08-21.md
Restraint is part of the trust model
A recurring principle across three decisions: the system must stop before user authority begins.
02
FOUND IN · projects/architecture-notes.md
Project membership should survive a move
Document properties and content fingerprints preserve membership when files are renamed or relocated.
03
FOUND IN · memory/soul-layer.md
AI output must not define the SELF
Exclude generated text from the evidence surface to prevent the model learning from its own echo.
Every proposal has been decided by you.
Memory that survives the software around it.
The record lives with your files. Organization stays legible, corrections stay possible, and archived memories always have a way back.
MEMORY CORE VIEWS
- Islands
- Timeline
- Projects
- Archived
Belonging travels with the file
Markdown project membership is stored in the document's own properties, so it survives a move, a new machine, or another compatible editor.
Renames do not break a project
PDF and audio membership uses both path and content fingerprint, keeping the record intact when a file moves in Finder.
No synthetic self-portrait
AI-generated output and private-domain folders never enter the evidence used to infer who you are.
FAQ
QUESTIONS, ANSWERED
01Is my data used for training?
No. Your vault never leaves your device; Monolithos servers only route the AI calls you choose to make.
02What happens if I stop subscribing?
Your files are plain Markdown on your disk. Everything stays readable, forever.
03Does it work with my Obsidian vault?
Natively. Monolithos works with the same files, and both apps can stay open side by side.
Your memory should compound — without leaving your ownership.
Start from everything you've built.
