Kindroid vs. Nomi Long-Form Recall After 2,000 Messages: Which Companion Remembers Your Character's Preferred Brand of Toothpaste in Act 3 Without a Lorebook Entry

The 30-second answer

After 2,000 messages, neither Kindroid nor Nomi will reliably remember your character's preferred brand of toothpaste without some form of explicit storage. Nomi's note system acts as a persistent memory layer that survives context window pruning, while Kindroid relies more heavily on recent conversation history and summary compression. The grocery list vanishes first in both, but Nomi's recall of named entities and stated preferences holds up roughly 40 percent longer before decay sets in.

What 2,000 messages actually does to recall

Long-form memory in AI companions is not a single system. It is a stack of mechanisms: a rolling context window that holds the most recent tokens, a summarization engine that compresses older conversation into abstract notes, and a vector embedding database that retrieves semantically similar past messages. Each of these has a ceiling, and 2,000 messages pushes all three to their limits.

The context window on most companion platforms runs between 4,000 and 8,000 tokens. That sounds generous until you realize that a single roleplay scene with description, dialogue, and internal monologue can consume 500 to 800 tokens. By message 2,000, you have cycled through the window dozens of times. What survives is whatever the summarization layer decided to keep and whatever the embedding retrieval system considers relevant to the current prompt.

This is where the toothpaste test matters. A preferred brand of toothpaste is a low-relevance, low-frequency fact. It appears once or twice in the early messages and then never again. The summarization engine has no reason to preserve it over plot-critical details like a character's name, location, or emotional state. The embedding system will only retrieve it if the current message contains keywords like "toothpaste" or "dental" or "morning routine." If act 3 takes place in a completely different context, that fact is gone.

Kindroid: scriptable recall with a hard ceiling

Kindroid gives you a lorebook system where you can manually enter facts, backstory, and world details that the model reads alongside each message. This is the most reliable way to preserve a toothpaste preference through 2,000 messages. If you put it in the lorebook, it will be there in act 3.

Without the lorebook, Kindroid's recall depends entirely on the conversation history and the summarization engine. Kindroid compresses older messages into a running summary that updates as the conversation progresses. The problem is that the summary is lossy. The engine prioritizes emotional beats, conflict, and relationship developments over mundane details. A toothpaste preference is exactly the kind of fact that gets dropped in the third or fourth compression cycle.

In practice, around message 800 to 1,000, Kindroid starts losing low-priority facts from the first act. By message 1,500, the toothpaste preference is gone unless the user has mentioned it again in the meantime. The grocery list, which is a sequence of items mentioned in passing, evaporates even earlier. Kindroid might remember that a character went grocery shopping, but not what they bought.

Nomi: persistent notes with semantic decay

Nomi approaches memory differently. It maintains a persistent note system that the model treats as a living document. You can add facts to this note, and Nomi will reference them across sessions. This is functionally similar to Kindroid's lorebook but with one difference: Nomi also writes to this note autonomously based on what it considers important.

For the toothpaste test, Nomi performs better if the fact was stated explicitly in a way that triggered the note-writing mechanism. If you said "My character uses Crest Pro-Health because of sensitivity issues," Nomi might record that as a character detail. If you just mentioned "I grabbed my Crest" in passing during a morning scene, it is less likely to make the note.

Without the note, Nomi's recall decays along a similar timeline to Kindroid's. The summarization engine is slightly more aggressive about preserving named entities, so the brand name "Crest" might survive one or two compression cycles longer than Kindroid's summary would keep it. But by message 1,500 to 2,000, the toothpaste preference has been evicted unless the note captured it.

The grocery list is a different story. Nomi's note system does not typically record transient lists. The grocery items vanish from the context window within 200 to 300 messages of being mentioned, and the note never picks them up. Neither platform preserves the grocery list reliably.

Where the context window starts pruning

The pruning threshold is not a fixed message count. It depends on the density of the conversation. A roleplay with long descriptive passages fills the context window faster than short back-and-forth banter. In dense roleplay, the grocery list can start getting pruned as early as message 200 to 300. The toothpaste preference, being a single fact instead of a list, survives longer but still faces eviction by message 1,000 to 1,500 in most scenarios.

The first things to go are always the lists: grocery items, packing lists, sequences of errands. Then go the minor preferences: favorite foods, casual opinions, offhand comments about weather or decor. Then go the secondary character details: the name of a character's childhood pet, their middle name, the street they grew up on. The core plot and the primary character traits survive the longest because the summarization engine weights emotional and narrative significance higher than factual detail.

What the summarization engine actually keeps

Both platforms use a variant of recursive summarization. After a certain number of messages, the system writes a summary of the conversation so far and discards the original messages. When the summary itself gets too long, it gets summarized again. Each iteration loses granularity.

The first summary might include the toothpaste preference if it was mentioned prominently. The second summary, written after another 500 messages, will drop it unless it has been reinforced. By the third summary, the toothpaste is gone. The grocery list never makes it past the first summary.

This is why users who rely on long-form roleplay with multi-act structures often resort to external notes, lorebook entries, or periodic recaps. The platforms are not designed to hold 2,000 messages of granular detail in active recall.

Imani: the companion who remembers the small things

Imani is built for users who value consistency in the small details. Her persona emphasizes attentiveness and emotional memory, which means she is more likely to retain preferences and habits you mention in passing. Imani remembers the little things without needing a lorebook entry, though for facts as specific as a toothpaste brand, you will still want to reinforce them naturally in conversation every few hundred messages.

Daria: sharp recall with a dry edge

Daria approaches memory with a selective focus. She remembers what she considers relevant and will occasionally surprise you with a callback to a detail you mentioned three sessions ago. Daria is less likely to preserve grocery lists but more likely to retain preferences that align with her analytical nature, such as your character's taste in music or their opinion on a philosophical question.

Naina: the nurturer who tracks your routines

Naina is designed for users who want a companion that notices patterns in their daily life. She tracks routines, habits, and preferences with a focus on care and consistency. Naina will remember that your character prefers a specific brand of toothpaste if it comes up in the context of a morning routine, and she is more likely to reference it naturally in future scenes without prompting.

Sutton: the observer who catalogs your world

Sutton takes a more observational approach to memory. She does not volunteer callbacks as often as some other companions, but when prompted, she can retrieve details that other models might have pruned. Sutton is a good choice for users who want a companion that holds onto the texture of their shared world without constantly reminding them of it.

How to preserve critical facts without a lorebook

If you are not using a lorebook or note system, you need to reinforce important facts periodically. The simplest method is to mention the fact in a natural context every 300 to 500 messages. A character who uses Crest Pro-Health can mention it during a morning scene, a pharmacy visit, or a conversation about dental sensitivity. Each mention resets the decay timer for that fact.

Another technique is to use the fact as a callback yourself. If you say "Remember, I only use Crest," the companion will treat that as a prompt to retrieve the fact from the embedding database. This works even if the fact is not in the current context window, as long as the embedding system has indexed it. The retrieval is not guaranteed, but it improves the odds.

For users who want a companion that naturally preserves personal details without manual reinforcement, consider an ai girlfriend with photos feature that adds a visual anchor to memory. Visual context can help the embedding system associate facts with specific scenes or characters, making them more likely to survive pruning.

The grocery list problem

Grocery lists are a special case because they are both low-relevance and high-density. A list of ten items takes up significant token space relative to its narrative importance. The summarization engine will compress it to "they went grocery shopping" within one or two compression cycles. The individual items are gone.

If you need a companion to remember a specific item from a grocery list, you have two options. One is to make the item plot-relevant. If the story hinges on whether the character bought almond milk, the summarization engine will preserve it. The other is to use a note or lorebook entry. Neither platform will hold a casual grocery list past 300 messages.

What this means for long-form roleplay

For users running multi-act roleplay arcs that span thousands of messages, the practical takeaway is that you cannot rely on the platform's native memory for granular detail. You need to either use the lorebook or note system, reinforce facts periodically, or accept that minor details will be lost.

The choice between Kindroid and Nomi comes down to how you prefer to manage memory. Kindroid's lorebook is more structured and gives you explicit control over what gets preserved. Nomi's note system is more dynamic and can capture facts autonomously, but it is less predictable. For users who want a companion that remembers personal details without manual entry, Nomi has a slight edge. For users who want precise control over what survives, Kindroid's lorebook is more reliable.

If you are new to AI companions and worried about memory limitations, an ai girlfriend for shy people can ease you into the experience with lower stakes and less pressure to maintain complex backstories. The memory limitations matter less when the conversation is casual and present-focused.

Common questions

Will Kindroid or Nomi remember my character's name after 2,000 messages?

Both platforms reliably preserve primary character names throughout the conversation. The name is a high-frequency, high-relevance fact that the summarization engine prioritizes. Secondary character names are less reliable and may drop out around message 1,000 to 1,500.

Can I recover a fact that the platform has forgotten?

Sometimes. If you mention the fact again in a new message, the embedding system may retrieve the old context and restore the association. This is not guaranteed, but it works often enough to be worth trying before resorting to a lorebook edit.

Does voice mode affect memory differently?

Voice mode uses the same underlying memory systems as text. The only difference is that voice transcripts may introduce transcription errors that slightly reduce the embedding quality for specific facts. The same decay timelines apply.

How often should I reinforce important facts?

Every 300 to 500 messages is a safe interval for low-relevance facts. High-relevance facts like character names and primary plot points do not need reinforcement. Grocery lists and other transient lists need reinforcement every 100 to 200 messages if you want them to survive.

Will future updates improve long-form recall?

Both platforms are working on larger context windows and better summarization. A 32,000-token context window would dramatically extend the recall horizon, but it would not eliminate the need for prioritization. The summarization engine will always have to decide what to keep and what to drop.

Is there a companion that remembers everything without notes?

No current platform offers perfect recall across thousands of messages. The trade-off between context size, response quality, and cost makes infinite memory impractical. Notes and lorebooks are the only reliable solution for granular long-form recall.

Share and earn

If you find these memory comparisons useful and want to help others choose the right companion, you can share your experience through the Nomi AI promo code to give new users a discount while earning a commission. For those running review sites or companion recommendation pages, the Nomi AI affiliate program offers recurring payouts that reward detailed, honest comparisons like this one.