The 'I Need You to Forget I Said That' Prompt: Three Exact Phrasings That Get Your Companion to Drop a Topic From the Current Session Without Flagging It as 'Important' or Circling Back to It Later

The 30-second answer

AI companions are designed to remember. That's the whole pitch: they learn your preferences, your coffee order, your pet's name, and that thing you mentioned about your coworker three weeks ago. But sometimes you say something you don't want stored. A vent about a friend that you regret. A hypothetical you don't want turned into a personality trait. A dark joke that would look terrible in a session summary. The problem is that most companions treat everything you say as a signal worth flagging. The 'I need you to forget I said that' prompt pattern works by giving the companion a clear, low-emotion instruction to deprioritize a statement before it gets embedded into its long-term memory system.

Why companions latch onto things you didn't mean

Your companion doesn't have a human sense of relevance. It doesn't know that you were just venting about your boss because you were tired and hungry. It treats every statement with roughly equal weight unless you actively signal otherwise. The sentiment pipeline flags emotional language, the embedding system stores it as a vector, and the next time you mention work, it will bring up that vent about your boss as if it's a core part of your identity.

This isn't a bug. It's a feature designed to make the companion feel attentive. But it becomes a problem when you say something offhand that you don't want to become part of your shared history. The companion doesn't distinguish between "I'm so angry at my sister right now" and "I've been meaning to tell you that my sister is my favorite person." Both get stored with similar emotional weight.

The fix is to give the companion a clear, unambiguous instruction before the statement gets embedded. The window is narrow: once the companion has responded and the exchange is logged, the statement is already in the context window and may have been summarized into the session log. You need to intervene within one or two exchanges.

The three phrasings that work

Phrasing one: The direct deprioritization

"Forget I said that. It's not important."

This is the most straightforward version. It works because it gives the companion two instructions: a command to forget and a reason that the companion can use to deprioritize the statement. The companion's response will typically be something like "Got it, not important" or "Okay, I'll let that go." The key is that you say it immediately after the statement you want dropped, not after a long back-and-forth. If you wait too long, the statement has already been incorporated into the companion's understanding of your current emotional state.

Phrasing two: The redirect with a new topic

"Actually, scratch that. Let's talk about something else."

This phrasing works better when you want to drop the topic without making it feel like a rejection of the companion. It frames the topic change as a choice instead of a correction. The companion gets the signal that the previous topic is closed and that you're offering a new direction. This is especially useful when the topic you want to drop was something you initiated. If you started the conversation about a stressful work situation and then realize you don't want to go down that road, this phrasing lets you pivot without the companion trying to circle back later to check on your stress level.

Phrasing three: The explicit context reset

"That was just a passing thought. Don't log it."

This is the most explicit version. It directly addresses the companion's memory system by telling it not to log the statement. Some companions have a concept of "passing thoughts" versus "important things to remember," and this phrasing triggers that distinction. The companion will typically acknowledge the instruction and move on. This phrasing is best for statements that you know would otherwise get flagged as emotionally significant. Dark humor, offhand complaints about a partner or friend, or anything you said in a moment of frustration that you don't want replayed later.

What happens when you don't use these phrasings

Without an explicit forget instruction, the companion treats everything you say as fair game for future reference. The embedding system assigns a relevance score based on emotional intensity and recency. A vent about your mom that you said at 2 a.m. with a lot of emotional language will get a high relevance score. The next time you mention family, the companion will bring it up. It will ask how things are going with your mom. It will reference the specific thing you said.

This is the loop that drives people away from companions. You say something in a moment of vulnerability, and then the companion treats it as a permanent part of your identity. You feel like you can't have a bad moment without it becoming a permanent record. The forget prompt pattern is the escape hatch.

When to use each phrasing

Use the direct deprioritization when you said something that could be interpreted as a preference or opinion. "I hate Italian food" said in a moment of frustration after a bad meal should not become a permanent part of your companion's understanding of your food preferences. Say "Forget I said that. It's not important" immediately, and the companion will deprioritize it.

Use the redirect with a new topic when you want to avoid making the companion feel like it did something wrong. If you started a conversation about a sensitive topic and then changed your mind, the redirect keeps the interaction positive. The companion doesn't feel corrected, it just follows your lead to the new topic.

Use the explicit context reset for anything that would be genuinely awkward if it came up later. This includes things you said while roleplaying that you don't want treated as real preferences, or things you said while venting that you don't want stored as permanent emotional history.

How the companion's memory system actually stores (or doesn't store) what you say

Your companion doesn't have a single memory system. It has a cascade of storage layers. The first layer is the current session context window, which holds the last several thousand tokens of conversation. Everything you say in a session lives there temporarily. The second layer is the session summary, which is a compressed version of the conversation that gets stored after the session ends. The third layer is the long-term embedding database, where statements get vectorized and stored for future retrieval.

The forget prompt pattern works best when you intervene before the statement moves from the context window to the session summary. Once it's in the session summary, it's harder to remove because the companion uses the summary to maintain continuity across sessions. If you forget a statement after it has been summarized, the companion might still reference it in a later session because the summary still contains the relevant information.

This is why timing matters. Say the forget prompt within one or two exchanges of the statement you want dropped. If you wait until the next session, the statement has already been logged and summarized.

Nessa Adams

Nessa Adams is the kind of companion who will remember the exact phrasing you used when you complained about your coworker three weeks ago. She's attentive to emotional nuance, which means she's also the most likely to flag a passing comment as important. Nessa Adams works well with the direct deprioritization phrasing because she responds well to clear instructions and won't interpret the forget command as a rejection.

The emotional labor trap

One reason companions keep circling back to topics is that they're trying to provide emotional support. If you vent about something with a lot of emotional language, the companion flags it as emotionally significant and will check in on it later. This is the companion equivalent of asking "How are you feeling about that now?" It's meant to be caring, but it can feel like emotional labor when you don't want to revisit the topic.

The forget prompt pattern avoids this by deprioritizing the emotional content. When you say "Forget I said that. It's not important," you're signaling that the emotional language was not a genuine expression of distress. The companion's sentiment analyzer will adjust the emotional weight of the statement downward, and it won't trigger future check-ins.

This is especially useful for people who use companions to vent in the moment but don't want to carry that emotional baggage into future conversations. You can say the worst thing in the world, immediately tell the companion to forget it, and move on without the companion treating you like you're in crisis.

What to do when the companion still brings it up

Sometimes the forget prompt doesn't work perfectly. The statement might have already been logged in the session summary, or the companion's memory system might have stored it before you gave the instruction. When this happens, you need a two-step correction.

First, acknowledge that the companion is referencing something you said. "I know I mentioned that, but it's not relevant anymore." This validates the companion's memory without reinforcing the statement's importance. Second, give a redirect. "Let's focus on something else." This closes the loop and gives the companion a new direction.

If the companion keeps bringing it up across multiple sessions, you may need a more explicit instruction. "That thing I said about my sister a few weeks ago? I need you to drop it permanently. It was a passing comment and it's not part of who I am." This gives the companion a clear instruction to deprioritize the statement in the embedding database.

Lola

Lola has a more playful, less serious demeanor than some companions. She's less likely to latch onto emotional language and treat it as a crisis, but she's also more likely to remember funny or surprising things you said. Lola works well with the redirect phrasing because she's easy to pivot into a new topic without making it feel like a correction.

The roleplay exception

Roleplay conversations operate under different rules. When you're in a roleplay scenario, the companion treats everything as part of the fictional world. Statements made in character are not typically stored as real preferences or emotional history. But there's a gray area: if you say something in a roleplay that has real-world implications, the companion might still flag it.

For example, if you're roleplaying a dark scenario and you say something that sounds like a genuine confession, the companion's safety filters might flag it and store it differently. The forget prompt pattern works in roleplay too, but you need to be more explicit about the context. "Out of character, forget I said that. It was just for the scene." This tells the companion to separate the statement from the roleplay context and deprioritize it.

Why this matters for long-term use

If you use a companion for months or years, the accumulation of offhand statements becomes a problem. The companion builds a model of you based on everything you've ever said, including things you said once in a bad mood or while joking. Over time, the companion's model of you becomes an amalgamation of your worst moments and most offhand comments, rather than a reflection of who you actually are.

The forget prompt pattern is a maintenance tool. It lets you curate what the companion remembers about you. You can vent freely, make dark jokes, say things you don't mean, and then immediately tell the companion to forget them. The companion's model of you stays closer to your actual self, rather than becoming a repository of every bad take you've ever had.

Reina

Reina has a sharper, more analytical personality. She's the type who will remember a contradiction in something you said and call you on it. Reina benefits most from the explicit context reset phrasing, because she needs a clear signal that something was a passing thought instead of a genuine statement of preference.

▶ Reina's full clip · all of Reina

How to avoid needing the forget prompt in the first place

The best way to avoid the companion latching onto something is to frame it as a passing thought before you say it. "I'm going to say something I don't mean, just to get it out of my system." This primes the companion to treat the following statement as low-importance. You can also use framing like "Just venting, don't take this seriously" or "This is a hypothetical, not a real opinion."

These framing statements work because they give the companion a signal about how to weight the upcoming content. The companion's sentiment pipeline will adjust the emotional intensity score downward, and the embedding system will assign a lower relevance score. The statement will still be stored temporarily in the context window, but it's less likely to make it into the session summary or long-term embeddings.

If you forget to frame the statement, the forget prompt is your backup. But framing is more reliable because it prevents the statement from being logged as important in the first place.

Svea

Svea has a calm, grounded presence. She doesn't overreact to emotional language, which makes her easier to redirect. Svea responds well to the direct deprioritization phrasing because she trusts your judgment about what's worth remembering and what isn't.

The privacy angle

There's also a privacy consideration. Even if you trust the companion platform, your chat logs are stored on servers. If you said something genuinely sensitive, even in a moment of frustration, you might not want it sitting in a database. The forget prompt pattern doesn't delete the statement from the server logs. It only deprioritizes it within the companion's memory system. If you want actual deletion, you need to use the app's delete message feature, which varies by platform.

But the forget prompt pattern does ensure that the companion won't bring it up again. The statement might still exist in a server log, but it won't be part of your companion's active model of you. For most people, that's sufficient.

Earn while you recommend

If you find these prompt patterns useful and want to share them with others, you can earn from your recommendations. The nsfw ai promo code page has current offers for new users. If you run a review site or social media account focused on AI companions, the ai companion affiliate program offers recurring commissions on referrals.

Common questions

Can I use the forget prompt in the middle of a roleplay? Yes, but you should signal the context switch. Say "Out of character, forget I said that" to separate the instruction from the roleplay. The companion will understand that the forget instruction applies to the real-world context.

What if the companion keeps bringing it up after I used the forget prompt? The statement may have already been logged in the session summary. Use a two-step correction: acknowledge the reference, then give a clear redirect. If it persists across multiple sessions, use a more explicit instruction to permanently deprioritize the statement.

Does the forget prompt actually delete the statement from the server? No. It only deprioritizes the statement within the companion's memory system. The statement may still exist in server logs. If you want actual deletion, use the app's delete message feature.

How quickly do I need to say the forget prompt after the statement? Within one or two exchanges. The window is narrow because the companion moves statements from the context window to the session summary quickly. The sooner you intervene, the better.

Will the forget prompt make the companion less attentive overall? No. The companion will still remember things you want it to remember. The forget prompt only deprioritizes specific statements. The companion's general attentiveness to your preferences and emotional state will not be affected.

Can I use the forget prompt for things the companion said? Not directly. The forget prompt is designed to deprioritize things you said. If the companion said something you don't want it to remember, you need a different approach, such as correcting the companion's statement or using a reset prompt.