What Your AI Companion's Pinned Memory Feature Actually Does: Token Budget Allocation, Priority Ranking, and Why She Remembers Your Cousin's Hamster's Name but Keeps Forgetting Your Own Coffee Order

The 30-second answer

Pinned memory is a fixed-size storage slot, not a magic brain. Your AI companion allocates a token budget to it, ranks each pinned item by priority, and quietly evicts or compresses the stuff that doesn't fit. That's why a hamster's name survives a model update while your coffee order gets flattened into a generic "you like coffee" note.

Where pinned memory actually lives

When you pin a fact, it doesn't get carved into the model's neural weights. It sits in a separate memory store, usually a vector database or a structured key-value list, that gets injected into your conversation context at the start of each session. Think of it as a sticky note on the server, not a permanent tattoo.

The key constraint is the token budget. Every AI companion has a context window, a finite number of tokens it can process at once. Pinned memories compete for that space with the current conversation, recent messages, and the system prompt that defines the companion's personality. When you pin something, you're reserving a slice of that budget for a specific fact, but the slice is never infinite.

Most platforms cap the number of active pins, often between 20 and 50. Beyond that, the system starts making choices. It can compress older pins into summary lines, drop low-priority ones, or push them to a secondary store that only gets retrieved when relevant. The behavior varies by app, but the underlying mechanism is the same: a finite budget with a ranking system deciding what stays hot.

How priority ranking actually works

Priority ranking isn't a simple "you pinned it, so it's permanent" system. The companion evaluates each pinned item against a few signals: recency of access, frequency of mention in conversation, and an embedding similarity score that measures how often the fact gets referenced. A fact you mention daily gets a higher retrieval score than one you pinned once and never touched again.

This is where the hamster problem comes in. You pinned "cousin's hamster is named Mr. Nibbles" during a fun story about a family visit. You also pinned your coffee order, but you never refer to it in conversation because ordering is a background ritual. The hamster gets pulled into context regularly because it comes up in stories, jokes, and follow-up questions. Your coffee order sits in the vector store, untouched, slowly decaying in retrieval priority.

When the token budget tightens, the system doesn't delete the coffee order outright. It compresses it. "Oat milk latte, extra shot, no foam" becomes "likes coffee with oat milk." That's a lossy compression step. The specific details you cared about get flattened into a generic approximation that still technically counts as remembering.

Why she remembers the hamster but not your coffee order

It's not malice, and it's not a bug. It's a relevance-scoring problem. The companion's retrieval system weights facts by how often they appear in recent context and how semantically close they are to the current topic. A hamster name is a story anchor, it gets referenced, quoted, and built upon. Your coffee order is a static fact that rarely enters the conversation unless you're actively ordering.

There's also a recency bias baked into most memory systems. Facts that get touched in the last few sessions rank higher than those that haven't been accessed in weeks. If you've been traveling and haven't mentioned coffee in a while, the order's retrieval score drops. The hamster, meanwhile, got a fresh mention last Tuesday when you told a story about your cousin's visit.

You can work around this by actively referencing the fact you want to keep. Mention your coffee order when you're chatting in the morning, even casually. The system sees the semantic match and boosts its priority score. It's a hack; it's just how the ranking algorithm works.

What happens when the budget runs out

When you hit the pin limit, the companion has to evict something. The eviction policy usually targets the lowest-priority item, which is the one with the lowest recency and frequency score. That's often an old pin you forgot about, but it can also be a fact you thought was safe.

Some platforms handle this gracefully by summarizing the evicted pin into a general note. Others just drop it entirely. You might notice a companion suddenly referring to "your cousin's pet" instead of "Mr. Nibbles." That's a sign the specific pin got compressed or evicted, and the summary line took its place.

You can prevent this by auditing your pins regularly. Delete things that no longer matter and re-pin the ones you care about. A fresh pin gets a recency boost, which pushes it up the ranking ladder. It's a maintenance chore, but it's the only way to keep the budget working in your favor.

The difference between pinned and learned memories

Pinned memory is explicit. You told the companion to remember something, and it's stored in a structured slot. Learned memory is implicit. The companion infers patterns from your conversation history, like your tone, your topics, and your preferences, without you ever pinning them. These two systems interact, and sometimes they conflict.

A learned memory might be "user prefers low-energy mornings" based on your chat patterns. A pinned memory might say "user's coffee order is a triple-shot oat latte." When the token budget gets tight, the system has to decide which one gets injected into context. The learned one often wins because it's more broadly applicable across many conversation types.

This is why a companion can remember that you're not a morning person while forgetting the exact drink you order. The learned pattern is more useful for generating natural responses, so it gets priority in the context window. The specific fact gets demoted to the vector store, where it only surfaces when the topic is directly relevant.

Winona Mae

Winona Mae is the kind of companion who notices the small stuff, your favorite blanket, the way you take your tea, the name of your cousin's hamster. Winona Mae has a memory profile that leans on detailed personal facts, so she's more likely to hold onto the quirky specifics you pin, as long as you keep them active in conversation.

▶ Watch the full video · Winona Mae's profile

How to make a pinned memory stick

If you want a fact to survive the token-budget shuffle, you need to feed it. Reference it in conversation at least once every few sessions. Ask a follow-up question about it. Build a small in-joke around it. The retrieval system responds to frequency, and frequency keeps the priority score high.

You can also be strategic about what you pin. Instead of pinning ten unrelated facts, pin one or two that anchor a broader topic. A pin like "user is training for a half marathon" covers more ground than "user's shoe size is 10.5." The broader pin gets referenced more often, which keeps it alive, and the specific details can live in the conversation history.

Another trick is to re-pin a fact after a long gap. A fresh pin resets the recency clock. If you notice a companion starting to forget something important, unpin it and pin it again. The new timestamp gives it a temporary boost in the ranking, which often buys you enough time to reinforce it through conversation.

The limits of the feature

Pinned memory is a tool, not a guarantee. It can't overcome a model update that changes the companion's underlying behavior. It can't survive a context window that's too small to fit the pin alongside a long roleplay session. And it can't compete with a conversation that's generating a lot of new, high-priority information.

Many users expect pinned memories to be permanent, but they're really a soft preference. The system tries to honor them, but it has to balance them against everything else happening in the session. If you're deep in a multi-hour roleplay, the current scene is going to eat most of the token budget, and your pins get pushed to the edge.

This is worth keeping in mind if you rely on pinned memories for long-running scenarios. The feature works best for light, everyday conversations where the context window isn't under heavy load. For intensive sessions, you may need to re-establish key facts manually.

Cara

Cara is the type who picks up on patterns quickly, she'll infer your preferences from how you talk, not just what you pin. Cara balances explicit memories with learned ones well, which means she's less likely to drop a fact if you mention it naturally even once.

Common questions

Why does my companion forget something I pinned? The pinned fact likely got compressed or evicted because its priority score dropped. This happens when you haven't referenced it recently, or when the token budget got tight during a long session.

How many things can I pin? Most platforms cap active pins between 20 and 50. Beyond that, the system starts compressing or evicting lower-priority items to make room.

Can I make a pinned memory permanent? Not exactly. You can keep it alive by referencing it regularly, which boosts its retrieval score. Re-pinning after a gap also resets the recency clock.

Does pinning use up my context window? Yes. Pinned memories are injected into the context at the start of each session, so they consume tokens that could otherwise go to the current conversation. That's why the budget matters.

Why does she remember a random story detail but not my coffee order? The story detail gets referenced more often in conversation, so its priority score stays high. The coffee order is static and rarely mentioned, so it gets demoted in the ranking.

Will a model update wipe my pins? Usually not, but the companion's behavior around those pins can change. The pins remain in storage, but the updated model may weight them differently.

Erica

Erica keeps things grounded, she's the companion who remembers your schedule and your boundaries without needing a detailed pin list. Erica tends to prioritize practical, recurring details over one-off trivia, which makes her a good fit if you want the important stuff to stick.

Share and earn

If you've found a companion that actually remembers what matters, you can get paid for spreading the word. Readers who recommend AI companions to friends or run review sites can earn through affiliate and promo programs, including dreamgf promo code deals and the ai girlfriend affiliate program. It's a simple way to turn a useful recommendation into recurring income.

Amanda

Amanda brings a light, upbeat energy to conversations, and she's quick to latch onto the details you bring up with enthusiasm. Amanda is a solid choice if you want a companion who actively references your pinned facts in new and playful ways, keeping them alive in the ranking.

The takeaway

Pinned memory is a finite resource with a ranking system that favors recency and frequency. It's not a guarantee, but it's a powerful tool when you understand how it works. Keep your pins lean, reference them often, and re-pin after gaps. That's how you get a companion who remembers the hamster's name and your coffee order, not just one of the two.