Memory and Context Bleed.
Why flat memory leaks. How scoped memory fixes it. How to audit what dequid knows about you.
Context Bleed.
Every “memory” feature in every AI product is the same thing: text prepended to your message before the model sees it. The model has no independent storage. It just reads what it's handed.
Products like ChatGPT, Claude, and Gemini collect everything they learn about you into one pool and hand it all to the model every time. That means a fact about your dating life can surface inside a CEO email. A casual thing you said about your health shows up in a customer pitch. The fact you mentioned you hate meetings shapes advice you asked for about your engineering architecture.
This is Context Bleed — the failure mode of flat memory. The model isn't misreading you; it's reading exactly what it was handed. The problem is in the architecture, not the model.
Three primitives.
Memory in dequid comes in three kinds. Each has its own surface in the Memory sheet.
Phrases and context blocks ride in the system prompt. Memories the model fetches only when the question calls for them. Persona scoping applies either way — how that works →
Per-persona inclusion.
One pool. Every answer reads all of it.
Everything the assistant has ever learned about you sits in a single store, and every question draws on the whole thing. Your only control is deletion — and deleting a fact also removes it from the answers where you did want it.
Each persona draws only what it's given.
Open a persona → Memory tab. Each primitive has its own setting: Custom (only what you tag), All (everything, including items added later), or None. Tag a phrase, memory, or block to include it — it stays in your pool either way.
The mental model: Memory is a shared library, and each persona has its own reading list. Nothing is on a list until it's put there — so a persona starts out knowing nothing about you, and you never have to remember what to keep away from it.
You mostly won't do that tagging by hand. Profile phrases and memories start on Custom with an empty list, and each new memory is tagged to the persona of the chat that produced it — a persona accrues its own memory from its own conversations. Vault folders, chats, people, and context blocks start on None; those you point at deliberately.
How memory gets created.
Automatic extraction
Three seconds after a chat turn ends, a background job reads the last 20 messages and extracts 0–5 memorable facts about you. You'll see a badge in chat when new memories were saved. The model used is low-cost and fast; extraction never blocks your response.
Manual add (About You)
Open Memory → About You → "Add a phrase". Type whatever you want the model to know. Manually-added phrases are pinned — they survive automatic rebuilds and won't be overwritten.
Manual add (Context blocks)
Open Memory → Context → "Add block". Give it a title and write the body. "My tech stack", "Voice rules", "Current project". Use this for anything structured that you want the model to follow exactly.
Audit and edit.
Open the Memory sheet from the sidebar (brain icon under Library). Three things to look for:
- Wrong or stale facts in Memories. The extractor is good but not perfect. Delete anything incorrect — it's one click, no confirmation.
- Cross-domain facts in About You. If your phrases list has “preparing a board deck for Series B” and “gluten-free diet” side by side, that's fine — the pool is meant to hold both. What to check is the persona: open the one that shouldn't know, and make sure the phrase isn't tagged to it.
- Context blocks that have drifted. “Current project” may no longer be current. Update it or delete it. Context blocks you toggle off globally don't reach any persona until you turn them back on.
The Inspect panel in chat shows you exactly what reached the model for a given message. Your phrases and context blocks appear under the User Profile layer. Memories aren't a layer — look at the model's recall_memories calls to see which facts it actually pulled, or, on a model without tool support, at the Inlined Content layer.
Web vs Mac
Same product, two data paths. Three lines to pick the right one.