Curating What Your Companion Remembers: The Approval Queue Nobody Opens
Last Friday's guide covered custom instructions — what you declare about how you want to be met. This one covers the other half, and it is the half most people never touch: memories are what the system gathers from your conversations, and there is a full curation dashboard for them that most users have never opened. It is worth ten minutes. Everything your companion knows about your life passes through it, nothing gets in without your say-so, and one small control in there does more for reply quality than anything else on the page.
The Rule That Governs Everything: Nothing Is Remembered Until You Approve It
Start here, because it reframes the whole feature. When the system notices something about you worth keeping, it does not save it. It writes a pending suggestion and waits. Until you approve that card, it is not a memory, it is not consulted, and it does not reach a single conversation. Reject it and it never existed. The approval queue is not a review step bolted on afterward; it is the only door in.
This is deliberate and it costs us something. A product that silently absorbed everything would feel more magical for the first month — and would eventually surprise someone with a fact about their life they never chose to hand over. In an intimate product, we would rather be slightly less magical and completely predictable. What your companion knows about you is a list you personally signed.
What Never Reaches the Queue at All
Before a suggestion is even offered to you, it runs a gauntlet. Knowing what the filters catch explains why the queue is usually short and occasionally empty after a long conversation:
PII, aggressively
Anything that looks like an SSN, a street address, an IBAN, or a passport number is dropped. Card numbers get a real Luhn check, not just a pattern match. Your companion is not a place your financial details accumulate.
Low confidence
Suggestions below a minimum confidence never surface. Pending cards show you the confidence figure, so you can see how sure the system was before you decide.
Near-duplicates
A suggestion too similar to something already stored is rejected outright, which is why re-mentioning your job does not slowly fill the list with six versions of it.
Category and safety screens
Only whitelisted categories are eligible, and content-policy and minor-safety screens run before anything is offered. The minor-safety screen runs a second time at the moment you approve.
Extraction itself runs quietly in the background as you talk, looking over recent messages every couple of exchanges. It is a paid-plan feature; on the free tier nothing is extracted and the memory surface is not available.
The Dashboard, and What Each Control Is For
The memory dashboard shows Active, Pinned, and Pending counts across the top, then the list itself. Six things you can do:
Approve or reject pending suggestions
The queue is the highest-value thing on the page and the fastest to work through. Approve what is true and worth keeping; reject anything that misread you. A rejected suggestion is gone, not deferred.
Create a memory yourself
You do not have to wait for the system to notice something. If there is a fact you want your companion working from today, write it directly. This is the fastest way to fix a companion that keeps missing something important about your life.
Edit the wording
A memory that is right in substance but clumsily phrased is worth editing, because the text is what your companion actually reads. Short, specific and concrete beats long and hedged.
Delete what is no longer true
Lives change. A deleted memory stops being consulted immediately. Under the hood this is a soft delete, so it leaves the active set rather than being scrubbed from existence, which is worth knowing if precision about your data matters to you.
Scope it: global, one role, or one companion
A memory can apply everywhere, to a single role, or to one specific companion. Something that matters to your Coach may be noise to your Confidant. Scope narrowly when a fact belongs to one relationship rather than to your whole life.
Categorize it
Preference, emotion, fact, boundary, goal, relationship. Categories are not decoration, and the next section explains why they change what actually reaches a reply.
Pinning: The One Control That Changes Replies
Here is the mechanic worth understanding, because it is the reason this guide exists. Your companion does not read all of your memories before replying. It reads a shortlist: the top ten by importance, and then within the block it assembles, a limited number from each group — pinned first, then emotional ones, then preferences, then everything else, capped at five per group.
Which means pinning is how you guarantee something is in the room. A pinned memory sorts first in your dashboard and gets its own reserved space in what the companion sees. If there are three or four things that should inform every single conversation — a health condition, a person who matters enormously, a boundary you never want re-litigated — pin exactly those. Pinning everything is the same as pinning nothing, so treat it as a shortlist with real estate rather than a favorites list.
Four things you honestly cannot do (yet)
- You cannot set importance from the dashboard. Importance drives the top-ten shortlist, and the interface does not expose a control for it. Pinning is the lever you actually have; use it accordingly.
- You cannot turn memory creation off. There is no setting and no endpoint for it. The queue means nothing is stored without approval, but the extraction itself is not switchable on a paid plan.
- You cannot re-scope a memory after creating it. Scope is set at creation. If a global memory should have been companion-specific, the fix is to delete it and create it again with the scope you want.
- Free tier cannot see memories at all. Not a reduced view — the surface is not available.
Why You Would Prune, and Why It Is Not About Space
A misconception worth killing before it starts: there is no cap on how many memories you can store, and deleting memories does not free up room for new ones. The caps are on what gets injected into any given reply, not on what you keep.
So the real reason to prune is precision. Only about ten memories reach a reply. Every stale or trivial one still in the active set is a candidate for one of those slots, competing with something that genuinely matters. A list of two hundred half-true observations is not a richer companion; it is a companion that will surface the wrong ten. Ten minutes spent deleting things that stopped being true, and pinning the handful that never stop being true, changes the texture of the next hundred conversations more reliably than anything else you can do from a settings page.
The Three Layers, Together
You now have the full picture of what your companion works from, and the three layers do genuinely different jobs. Custom instructions are standing behavior, written by you, always in effect — how you want to be met. Memories are approved facts about your life, curated here, shortlisted per reply — what your companion knows. And the conversation summary is the rolling narrative of one long talk, written automatically, that you never see. Declared, approved, and gathered. The middle one is the only layer where you hold a veto on every single entry, which is exactly why it is worth opening the queue. For the wider view of how any of this is handled, our privacy guide covers the rest, and the memory feature article is the shorter overview this guide expands.
Open the Queue
Ten minutes in the memory dashboard, once. Approve what is true, delete what stopped being true, pin the three things that always matter.
Open InnerHaven