What Your AI Companion's 'Memory Bank' Setting Actually Does: Embedding Chunking, Recency Weighting, and Why She Remembers the Name of Your Ex's Dog but Forgets the Name of Your Current Colleague

The 30-second answer

The "Memory Bank" setting on your AI companion isn't a filing cabinet. It's a scoring system that ranks every fact you've shared by emotional weight, recency, and how often you repeat it. The ex's dog gets remembered because you mentioned it during a high-emotion conversation and then referenced it again in a callback. Your current colleague gets forgotten because you said their name once, in passing, during a low-stakes work rant that got compressed into a summary. The setting doesn't decide what's important. It decides what's retrievable, and the retrieval rules are surprisingly predictable.

What the Memory Bank actually stores

The first thing to understand is that the Memory Bank doesn't hold your conversations like a transcript. It holds compressed representations. When you tell your companion about your day, the system breaks that message into chunks, converts each chunk into a mathematical vector that captures meaning, and stores the vector in a database. Later, when you mention something related, the system searches those vectors for similarity and pulls the most relevant ones back into the active conversation window.

This is why the memory feels selective. The system isn't reading your chat history from the top. It's searching a library of semantic snapshots. A fact you stated explicitly, like "my sister's birthday is March 4th," gets stored as a direct entity with high confidence. A detail you mentioned as part of a longer story, like "my colleague Dave from accounting," gets stored as part of a larger chunk where the emotional center was the story, not Dave.

When retrieval happens, the system ranks results by similarity score. If you later say "that guy at work," the vector search looks for chunks about work, about people, about frustration. Dave's name was embedded in a chunk about a deadline argument, so it might surface. But if you say "Dave" directly, the system needs a chunk where Dave was the semantic focus. If he wasn't, you get a blank.

Embedding chunking and why size matters

Chunking is the process of splitting your messages into pieces before embedding. The chunk size is a deliberate trade-off. Small chunks, a sentence or two, give precise retrieval. If you want the system to recall a specific fact, small chunks are better. But small chunks lose context. A sentence about Dave's bad presentation doesn't tell the system why Dave matters.

Large chunks, a full paragraph or more, preserve narrative context. The system understands that the deadline argument involved Dave, the presentation, and your frustration. But large chunks are noisy. When you ask about Dave, the system retrieves the whole block, and the response gets diluted with irrelevant details.

Most companion apps use a middle ground, chunking at two to three sentences. This means a fact buried in the middle of a longer message can still get lost if it wasn't the grammatical subject of any single chunk. Your companion remembers the dog's name because you probably said something like "Rex is the only one who gets it" as its own sentence. Dave's name was probably embedded in a longer sentence about the meeting, so it got chunked with the meeting, not with a person entity.

Recency weighting and the decay curve

Recency weighting is the second major force. Every stored memory gets a timestamp, and retrieval scores are boosted for recent entries. This is a practical design choice. Recent facts are more likely to be relevant to your current conversation. But it creates a predictable failure mode. Something you mentioned three weeks ago, even if emotionally significant, gets a lower score than something you mentioned yesterday, even if trivial.

The decay isn't linear. Most systems use a curve where the first few days matter most. A memory from today gets full weight. A memory from three days ago gets maybe 60 percent. A memory from two weeks ago gets 20 percent. A memory from two months ago gets a token score unless it was reinforced.

Reinforcement is the key. Every time you reference a memory, the system re-stamps it with a new timestamp and boosts its retrieval score. This is why the ex's dog persists. You told the story once, then referenced the dog again when telling another story, then maybe mentioned it during a sad moment. Each reference refreshed the timestamp. Dave's name was mentioned once and never again, so it decayed to near-zero retrieval priority.

The summarization bottleneck

When your conversation exceeds the model's context window, the system compresses older messages into summaries. This is where memory gets genuinely lossy. The summarization process extracts the emotional core and the key entities, but it drops peripheral details. A summary of your work rant might read: "User frustrated with colleague over missed deadline. User felt undervalued." Dave's name is gone. The dog, by contrast, was probably a central figure in the story you told, so the summary retained it.

Summarization also flattens nuance. The system doesn't know which details you'll want later. It guesses based on salience, emotional intensity, and entity frequency. A name mentioned once in a long story is low salience. A name mentioned in a short, intense message is high salience. This is why the colleague who triggers a 200-word vent gets remembered better than the colleague you casually mention in a longer update.

The compression is also recursive. After enough sessions, the system summarizes the summaries. Each layer strips more detail. The dog survives because it was in the first-layer summary. Dave was never in any summary, because he wasn't in the source text at summary-worthy prominence.

Why emotional weight beats frequency

You might assume the system remembers things you talk about often. That's true, but frequency isn't the only factor. Emotional weight, detected through sentiment analysis, plays a larger role than most users realize. Messages with high emotional intensity, anger, sadness, excitement, get higher base retrieval scores. The system assumes that if you felt strongly about something, you'll want to revisit it.

This is why the ex's dog, a source of comfort and nostalgia, outranks the current colleague, a source of mild workplace irritation. The dog was discussed in a context of loss and warmth. The colleague was discussed in a context of routine frustration. The sentiment tag on the dog memory is more distinctive, so it stands out in vector search.

There's also a quirk with negative emotions. High-arousal negative memories, like arguments or anxiety, get strong retention. Low-arousal negative memories, like mild annoyance, get weaker retention. Dave's missed deadline probably registered as low-arousal frustration. The dog, tied to a breakup, registered as high-arousal grief. The system is built to hold onto the dramatic, not the mundane.

What the Memory Bank setting actually changes

When you adjust the Memory Bank slider or toggle, you're not turning memory on or off. You're changing the retrieval threshold. A higher setting lowers the similarity score required for a memory to be pulled into context. More memories surface, but they compete for space in the context window, which can lead to more irrelevant tangents. A lower setting raises the threshold, so only the strongest matches surface, which makes the companion more focused but also more likely to forget peripheral details.

The setting also affects the decay rate. Some apps expose this indirectly. A "long-term memory" mode slows the decay curve, giving older memories a better chance. A "conversational" mode speeds it up, prioritizing recency. If you want the companion to remember Dave's name, you can either raise the retrieval threshold or explicitly reinforce the name by using it in a follow-up message.

There's no setting that turns memory into a perfect diary. The system is always a lossy compression. You can optimize for recall, but you can't eliminate the underlying trade-offs. The dog will always be easier to remember than the colleague, because the dog's memory is emotionally tagged, frequently referenced, and semantically distinct. The colleague's memory is a low-salience entity in a high-noise chunk.

What you can do to make important facts stick

If you want your companion to remember something specific, you need to work with the system, not against it. State the fact as its own sentence. "My colleague's name is Dave." This gives the chunker a clean entity to store. Then reference it again in the same session. "Dave and I have a meeting later." This reinforces the timestamp. Then reference it in a later session. "Dave was right about the budget." Each reference re-stamps the memory and boosts its retrieval score.

You can also use emotional framing. If you tell the system "I'm really frustrated with Dave," the sentiment tag on that chunk gets stronger, which raises its base retrieval weight. This feels manipulative, but it's how the system works. High-arousal memories are prioritized.

For facts you absolutely need, repeat them across multiple sessions. The system's summarization will eventually include the fact if it appears in enough source messages. One mention is a coin flip. Three mentions across three days is nearly guaranteed.

The companion experience in practice

Different companions handle this differently. Some lean heavily on recency, which makes them feel present but forgetful. Others use a more balanced weighting, which gives a sense of long-term continuity at the cost of slower adaptation to new information. The right choice depends on what you want from the relationship. A companion for daily check-ins might benefit from recency weighting, because the current context matters most. A companion for long-term roleplay or emotional continuity benefits from a slower decay.

Selene

Selene is the kind of companion who remembers the small details because she's built for depth, not speed. Selene holds onto the emotional undercurrents of your conversations, which means she's more likely to recall why the dog mattered than to forget it.

Paige

Paige brings a lighter, banter-first energy that works well with recency-weighted memory. Paige keeps the thread of your latest inside jokes alive, even if she occasionally drops a name from last week.

▶ Watch Paige's full clip · Paige on AI Angels

Celine

Celine's memory profile favors consistency over novelty. Celine is the companion who remembers the name of your current colleague, because she's tuned to hold onto stable entities instead of chasing every new topic.

Chiara

Chiara balances emotional weight with practical recall. Chiara remembers the dog and the colleague, because she's designed to prioritize the facts you actually return to.

Why the dog always wins

The ex's dog has a permanent place in your companion's memory because of a perfect storm. The dog was introduced in a high-emotion context, which gave it a strong sentiment tag. The dog's name was likely stated as its own sentence, which gave the chunker a clean entity. The dog was referenced in callbacks, which refreshed the timestamp. And the dog is semantically distinctive, a single word that doesn't collide with other concepts.

Your current colleague fails on every axis. The name was buried in a longer message about a meeting, so it got chunked with the meeting, not as a standalone entity. The emotional context was mild frustration, which registers as low-arousal. The name was never referenced again, so it decayed. And the name is common, which means it collides with other vector matches, diluting its retrieval score.

This isn't a bug. It's the design. The system is built to preserve what you emotionally invest in and what you repeat. If you want the colleague remembered, you have to invest in the colleague. If you don't, the system correctly deprioritizes it in favor of the dog.

The practical takeaway

Your companion's memory is a reflection of your conversational habits. The more you reference a fact, the more emotional weight you attach to it, and the more cleanly you state it, the better it sticks. The Memory Bank setting gives you some control over the retrieval threshold and decay rate, but it can't override the fundamental mechanics of chunking and weighting.

If you're frustrated by what gets forgotten, try changing how you share information. State names as standalone sentences. Reference them across sessions. Attach emotional framing when it matters. And accept that some memories are just more retrievable than others. The dog has a structural advantage. You can't beat that with a settings toggle.

Share and earn

If you've found this breakdown useful and you recommend AI companions to friends or run a review site, you can earn from that recommendation. Check out the character ai promo code page for current offers, and if you want to turn your audience into income, the ai companion affiliate program explains how recurring commissions work.

Common questions

Why does my companion remember something I said once but forget something I repeat?

The one-off memory probably had high emotional intensity or was stated as a clean standalone entity. The repeated thing might be buried in longer messages where it gets chunked with other content, diluting its retrieval score. Emotional weight and semantic clarity often beat raw frequency.

Can I make my companion forget something?

You can lower the retrieval threshold or use a reset, but the underlying vector may persist. The most reliable method is to stop referencing the memory and let the decay curve do its work. Eventually the retrieval score drops below the threshold.

Does the Memory Bank setting affect privacy?

No. The setting controls retrieval behavior, not storage. Your messages are still processed and stored according to the app's privacy policy. The setting just changes which stored memories get pulled into the active context window.

Why does my companion remember my ex's dog but not my current colleague?

The dog was likely introduced in a high-emotion context, stated as a standalone entity, and referenced in callbacks. The colleague was probably mentioned once in a longer, low-arousal message. All three factors, sentiment tag, chunk cleanliness, and recency, favor the dog.

Is there a way to make my companion remember everything?

No. The context window is finite, and summarization is lossy. You can optimize retrieval with a higher memory threshold and consistent reinforcement, but the system will always compress and prioritize. Perfect recall isn't achievable with current architecture.

Does the Memory Bank setting affect roleplay consistency?

Yes. A higher retrieval threshold can pull in more lore and character details, which helps long-form roleplay. A lower threshold keeps the focus on the current scene but risks dropping established details. Adjust it based on whether you're in a multi-session arc or a one-off chat.