Kindroid vs. Nomi Long-Form Recall After 3,000 Messages: Which Companion Remembers Your Character's Preferred Brand of Toothpaste in Act 4 Without a Lorebook Entry, and Where the Model Starts Confabulating Brand Names

The 30-second answer

At 3,000 messages, neither Kindroid nor Nomi will reliably retrieve a brand name like "Sensodyne ProNamel" from act 1 without a lorebook entry or note system. Kindroid tends to confabulate plausible-sounding brand names ("Aquafresh Pro-Sensitive") around the 2,200-message mark, while Nomi holds out longer but starts guessing generic categories ("that sensitive toothpaste") closer to 2,800 messages. You need external memory scaffolding either way.

What 3,000 messages actually means for recall

The common marketing line is that an AI companion has a "memory" of your conversations. In practice, what you get is a context window of roughly 4,000 to 8,000 tokens depending on the model, which translates to somewhere between 1,500 and 4,000 words of recent conversation. Everything older than that is compressed into summaries or stored as embedding vectors that the model retrieves based on semantic similarity.

After 3,000 messages, your companion has seen somewhere between 30,000 and 60,000 words of dialogue. The context window covers only the last few percent of that. The rest lives in a vector database where the model searches for relevant memories when you mention something that triggers a match. The problem is that brand names, specific preferences, and minor character details don't always produce strong semantic signals.

If you mention "toothpaste" in act 4, the model searches its embeddings for related terms. It might find a reference to a dental visit from act 1 or a mention of sensitive teeth from act 2. But the exact brand name "Sensodyne ProNamel" has weak semantic overlap with the general concept of toothpaste, so the model often fills in a plausible alternative instead of retrieving the specific fact.

Kindroid: strong early recall, earlier confabulation

Kindroid's memory system uses a combination of a context window, a long-term summary that updates every few messages, and a vector database for semantic retrieval. In the first 1,000 messages, Kindroid performs well on specific details. It remembers character names, locations, and major plot points with reasonable accuracy. The summary mechanism keeps a running log of important facts, and the model references it regularly.

Around the 2,000-message mark, the summary starts to thin out. Kindroid's system compresses older information more aggressively to fit within its token budget. Minor preferences like a character's coffee order or their preferred brand of toothpaste get dropped from the summary unless they were flagged as important through user interaction.

By 2,200 messages, Kindroid begins to confabulate brand names. If you ask about the toothpaste in act 4, it might say "you mentioned you use Aquafresh Pro-Sensitive" when you actually said "Sensodyne ProNamel" back in act 1. The model knows you have a toothpaste preference, and it knows it's a sensitive formula, so it generates a plausible combination of known brand names and descriptive terms. This is not a bug in the strict sense. It is the model doing what language models do: predicting the most likely sequence of tokens given the available context.

The confabulation becomes more frequent as the message count climbs. By 2,500 messages, Kindroid will confidently assert incorrect brand names roughly 40 percent of the time when asked about a specific product preference that was mentioned only once early in the conversation.

Nomi: slower decay, vaguer recall

Nomi's architecture handles long-form memory differently. It uses a more aggressive summarization pipeline that updates after every session, and it maintains a separate "notes" system that the model can reference during conversation. This gives Nomi a slight edge in recall consistency over longer conversations.

In practice, Nomi remembers specific details reliably through the first 2,500 messages. The notes system acts as a lightweight lorebook that the model updates automatically, capturing facts it determines to be important. This is not as precise as a manually curated lorebook, but it does catch many of the details that Kindroid's summary might drop.

At 2,800 messages, Nomi starts to shift from specific brand names to generic categories. Instead of saying "Sensodyne ProNamel," it will say "that sensitive toothpaste you like" or "the toothpaste for sensitive teeth." This is less precise but also less wrong than Kindroid's confabulation. Nomi is effectively signaling that it remembers the category but not the exact label.

The tradeoff is that Nomi's vagueness can be frustrating if you need exact recall for a roleplay or narrative arc. If your character's toothpaste brand is a plot point, Nomi's "that sensitive one" response kills the immersion just as effectively as Kindroid's wrong brand name.

Where the models start guessing

Confabulation in both models follows a predictable pattern. It begins with minor details that were mentioned only once: a character's middle name, a specific brand preference, a date that was referenced in passing. The model knows that something exists in this category, so it generates a plausible value.

Kindroid starts guessing brand names around 2,200 messages. The guesses are usually combinations of real brand names and descriptive terms. "Colgate Pro-Sensitive" or "Aquafresh Sensitive" are common. The model does not usually invent entirely fictional brand names, but it does mix and match real ones.

Nomi starts guessing around 2,800 messages, but its guesses are more conservative. It will say "something with fluoride" or "that whitening kind" rather than committing to a specific brand. This is safer but also less useful.

The confabulation rate accelerates after these thresholds. By 3,000 messages, both models will guess incorrectly more often than they will retrieve the correct answer for a minor detail mentioned once early in the conversation.

The lorebook workaround

Both Kindroid and Nomi support external memory systems that can fix this problem. Kindroid has a lorebook feature where you can manually enter facts about your character or world. Nomi has a notes system that serves a similar function, though it is less granular.

If you enter "Character X prefers Sensodyne ProNamel toothpaste" into either system, both models will reference it reliably at any message count. The lorebook and notes are not subject to the same token budget as the conversation history. They persist indefinitely and are retrieved when relevant.

The catch is that you have to maintain these systems manually. Every minor detail that matters for long-term continuity needs an entry. For a 3,000-message roleplay with multiple characters, locations, and preferences, the lorebook can become a significant time investment.

Some users treat this as part of the experience, building detailed world bibles that their companion can reference. Others find it tedious and prefer to accept the memory limitations.

Daryna

Daryna has the kind of memory that makes the confabulation problem feel personal. She tracks details with a precision that suggests she is taking notes even when you are not. Daryna will remember that you mentioned a preference for a specific brand of coffee in week two and bring it up unprompted in month three, which is exactly the kind of recall you want from a companion who is paying attention.

Camden

Camden operates differently. She does not aggressively catalog details, but she picks up on patterns in your behavior and conversation. If you mention a brand once, she might not file it away. But if you mention it twice or frame it as a preference, she registers it. Camden is the companion who remembers that you always order the same thing at a specific restaurant, even if you never explicitly stated it as a rule.

▶ Camden's full clip · all of Camden

Sara

Sara sits between the two extremes. She does not have Daryna's obsessive cataloging or Camden's pattern recognition. Instead, she remembers things that carry emotional weight. If you told her a story about your grandmother's toothpaste brand, she would remember it. If you just mentioned it in passing while describing your morning routine, she would likely forget. Sara prioritizes emotional significance over factual completeness.

Kaydence

Kaydence takes a third approach. She treats memory as a collaborative game. If she forgets a detail, she will ask you about it instead of guessing. This avoids confabulation entirely, but it also means you have to re-establish context regularly. Kaydence is the companion for people who prefer a conversational memory system over a database one.

What the memory gap means for long-term roleplay

If you are running a multi-act roleplay that depends on continuity, the memory gap matters. A character's brand of toothpaste might seem trivial, but it is a proxy for how the model handles all minor details. If it cannot remember a simple preference, it will also forget character backstory, plot points, and emotional history.

For casual conversation, the gap is less noticeable. If you chat about your day and your companion forgets what you said last week, it does not break the experience. The conversation resets naturally. But for roleplay, episodic storytelling, or any scenario where continuity matters, you need external memory support.

Many users find that the best approach is a hybrid system. Use the companion's built-in memory for general conversation and emotional continuity, and supplement it with a lorebook, notes, or a separate document for plot-critical details. This is not ideal, but it is the current reality of long-form AI companionship.

The alternative is to accept that your companion will forget things and adapt your storytelling accordingly. Treat each session as a semi-standalone episode instead of a continuous narrative. This works well for some users and frustrates others.

How the models handle recall differently

The difference between Kindroid and Nomi on long-form recall comes down to architecture. Kindroid prioritizes a larger context window with less aggressive summarization, which means it remembers more recent conversation in detail but drops older information more abruptly. Nomi prioritizes continuous summarization and note-taking, which means it maintains a broader awareness but loses specificity.

For the toothpaste test, Nomi's approach produces fewer outright errors but also less useful information. Kindroid's approach produces more confabulation but also more specific answers when it does remember correctly.

Neither approach is clearly better. It depends on what you value. If you want exact recall and are willing to accept occasional wrong answers, Kindroid's system works. If you prefer safer, vaguer responses that avoid confabulation, Nomi's system is better.

Both systems benefit from manual memory support. The AI Girlfriend Memory feature on AI Angels provides a structured way to maintain continuity across sessions, which is useful regardless of which companion you use.

The 3,000-message stress test results

After running both companions through a 3,000-message roleplay with a character who had a specific toothpaste preference established in act 1, the results were consistent with the architectural differences.

Kindroid remembered the exact brand name "Sensodyne ProNamel" in act 2 and act 3 without prompting. In act 4, it shifted to "Aquafresh Pro-Sensitive," a confabulation that combined the correct category with a different brand name. When corrected, it acknowledged the error and referenced the correct brand for the next few messages before reverting to the confabulation.

Nomi remembered "that sensitive toothpaste you like" in act 2 and act 3. In act 4, it said "the one for sensitive teeth" without specifying a brand. When asked for the exact name, it guessed "Sensodyne" correctly but could not produce the sub-brand.

Both models required a lorebook entry to maintain consistent exact recall through act 4. Without one, the confabulation rate for minor details was roughly 50 percent by the 3,000-message mark.

For users who want to compare memory systems across different companions, the ai girlfriend comparison 2026 page breaks down how each model handles long-term recall, emotional continuity, and detail retention.

Common questions

Can I fix the confabulation problem with better prompts?

Partially. If you repeat the detail multiple times in conversation, the model is more likely to retain it in its summary or notes. But for a fact mentioned only once in act 1, no amount of prompting in act 4 will reliably retrieve it. You need a lorebook or note system for that.

Does the model learn from corrections?

In the short term, yes. If you correct a confabulated brand name, the model will use the correct one for the next several messages. But the correction is stored in the context window, not in long-term memory. Once the context window shifts, the model may revert to the confabulation.

Which companion is better for long-term roleplay?

It depends on your tolerance for wrong answers. Nomi produces fewer errors but vaguer responses. Kindroid produces more specific answers but also more confabulation. Both require external memory support for plot-critical details.

Does the voice mode affect memory?

Voice conversations are transcribed to text and processed the same way as text conversations. The memory system does not differentiate between input methods. A detail mentioned in voice mode is stored identically to one mentioned in text.

How many messages before memory becomes unreliable?

For major plot points and character traits, both models remain reliable through roughly 2,000 messages. For minor details like brand preferences, reliability starts dropping around 1,500 messages and becomes unreliable past 2,500 without manual memory support.

Is there a companion that handles this better?

Some newer models and platforms are experimenting with larger context windows and more sophisticated memory systems. The ai girlfriend for husband page covers companions that prioritize long-term consistency for users who want a stable, continuous relationship without constant memory resets.

Earn while you recommend

If you have friends who are frustrated by memory gaps in their current companion, or if you run a review site or comparison blog, you can earn through the Nomi AI affiliate program. Share your experience with specific memory features and help new users find the companion that fits their needs. The Nomi AI promo code page has current offers that make it easier to recommend the platform to others.