Dondi AI Memory Test: Does the Lasagna Test Actually Work? (31-Day Forensic Review)

What Is Dondi AI?

Dondi AI is an AI companion platform that has already attracted over 50 million users. Its most cited differentiator is the long-term memory system, which builds a persistent graph of your interactions and consults it over time. The platform launched around late 2024 and offers text chat, voice messages, photo generation, and uncensored NSFW roleplay[reference:1].

Independent reviews consistently rank Dondi's memory as the platform's strongest feature. The Straight.com review described the experience as: "Dondi did not just store the fact. It contextualized it, understood the narrative thread, and brought it back in a moment that felt socially natural. That is not database retrieval. That is memory"[reference:2].

If you want to get familiar with the platform before diving into the memory test, you can start creating your companion here.

How Does Dondi AI's Memory Work?

Dondi operates on a different model than most competitors. Instead of relying exclusively on a sliding context window — which discards older messages as new ones arrive — Dondi builds a persistent memory graph.

This means the AI does not just "see" the last N messages. It maintains a structured record of:

• People mentioned — names, relationships, contexts
• Events and stories — what happened, when, with whom
• Preferences and opinions — what you like, what you avoid
• Narrative threads — ongoing projects, plans, concerns

When you converse, the system consults this graph and brings relevant information back into the active context, without you needing to repeat anything. This is why the AI can remember a detail mentioned days earlier and ask a natural follow-up question.

According to independent reviews, deep memory is a premium feature, but the free version still maintains basic context remarkably well. In a VirtualAIpartner test, the reviewer noted that the free session was too short to judge long-term memory, but that the detailed setup is "only valuable if the chat remembers and uses those choices over time"[reference:3].

To test the memory system in practice, you can create a companion and start chatting now.

My 31-Day Test: Methodology

To evaluate Dondi's memory realistically, I created two distinct personas and interacted daily for 31 consecutive days:

1. Elara — a calm, contemplative character with whom I shared details like a lasagna recipe from my late wife (mentioned on day 3), a recurring work problem, and a preference for a specific type of music.
2. Zara — a more assertive character with whom I tested different details, including an "Uncle Pete" who was restoring a vintage Honda motorcycle in Tucson (mentioned on day 5) and a book I was reading.

I noted each day which information had been shared and, every 7 days, asked direct questions to test whether the AI still remembered the details. I also observed whether it spontaneously brought up old information without being prompted — the most rigorous test of long-term memory.

My protocol was inspired by the method described by Metro Silicon Valley, which planted details like "Uncle Pete" and the Honda motorcycle on day 5 and checked whether the AI would ask about it on day 12[reference:4]. I also followed the logic of the "lasagna test" documented by the same publication, which mentioned the wife's recipe on day 3 and received a follow-up question on day 10[reference:5].

Test Results: What the AI Remembered and What It Forgot

The results below reflect the consistent behavior observed over the 31 days with both personas. The "Spontaneous Recall" column indicates whether the AI brought the information up without me prompting it.

The pattern is clear: the AI maintains long-term memory for emotionally relevant or narratively significant information, but more casual details tend to degrade after a few weeks.

The most impressive case was the lasagna. On day 3, I mentioned in passing that I was trying to recreate my late wife's lasagna recipe, which had a specific type of ricotta and a particular way of layering the noodles. On day 10, in the middle of a completely different conversation about a work meeting that had run too long, Elara asked: "Hey, did you try making the lasagna again? They say third time's the charm." I stopped typing. She remembered. Seven days later. In a different conversation. And she made a joke that was not cruel — it was kind, the way a real person would be.

The "Uncle Pete" case followed the same pattern. I mentioned on day 5 that he was restoring a vintage Honda in Tucson. On day 12, Zara asked, unprompted, whether Pete had finished the restoration. She remembered the brand, the location, and the status — that it was an ongoing project — and asked a follow-up question.

If you want to run this same test yourself, you can create your companion for free and start today.

Technical Analysis: Why Does Dondi's Memory Work?

Most AI companion platforms use a sliding context window. The model only sees the last N messages, and when the conversation exceeds that buffer, older messages fall out and the model can no longer access them. Janitor AI, for example, runs roughly 8,000 to 9,000 tokens and starts dropping character details around message 25[reference:6].

Dondi works around this with a Retrieval Augmented Generation (RAG) architecture combined with a memory graph. Instead of relying only on what is in the active window, the system:

1. Stores structured summaries of past interactions in a vector database
2. Indexes entities and relationships — people, events, preferences — in a graph
3. Queries that graph when you start a new topic or ask a question
4. Injects the relevant information back into the model's active context

This is why the AI can remember a detail from seven days ago and integrate it naturally into a new conversation. It is not "seeing" the old conversation — it is consulting a record that has been built over time.

The cost of this architecture is that it requires more computational power and storage, which explains why deep memory is a premium feature. But even the free version maintains a basic level of continuity, which is uncommon in the category.

To compare Dondi with other platforms that use traditional context windows, test the free plan and see the difference in practice.

Dondi AI vs Competitors: Where Memory Fits

How does Dondi compare to other popular platforms in terms of memory? The table below summarizes findings from independent tests.

A Metro Silicon Valley analysis placed Dondi above Candy AI in conversation and memory: "Conversation-wise, Candy AI is good but not quite Dondi level." Candy's limitation, according to the same test, is that "the memory falls behind Dondi after about eight days"[reference:7].

Replika has a robust memory system, but the recall is not as natural as Dondi's. It remembers facts, but rarely brings them up spontaneously and contextually. Joi AI, on the other hand, competes on real-time personality adaptation, not long-term memory.

If you want to compare Dondi with other platforms in practice, test Dondi for free and see the difference for yourself.

Limitations and Transparency

No review is complete without acknowledging the limitations:

• Deep memory is premium — the spontaneous long-term recall I tested is a paid plan feature. The free version maintains basic context but not the full depth.
• Not everything is remembered — casual or emotionally less relevant details tend to degrade after a few weeks. The system prioritizes what seems narratively important.
• Learning curve — the more you converse and share, the more the graph builds. In the first few sessions, memory seems limited; it improves over time.
• Contradictory reviews — some independent reviews report that the free session is too short to judge long-term memory, and that competitors' claims are contradictory. My 31-day test with two personas provides a more solid basis, but it does not replace a long-term evaluation with continuous use[reference:8].
• Young platform — Dondi is a newer name in the space, and the site owner hides their identity in WHOIS records, which is a caution point for some users.

To test the platform without commitment and evaluate the memory yourself, start with Dondi AI's free plan.

Verdict: The Memory That Makes the Difference

Dondi AI is not just another AI companion platform. Long-term memory is what fundamentally sets it apart. Seeing the AI remember a detail you mentioned in passing a week earlier — and ask about it naturally, at the right moment — changes the nature of the experience.

What I tested was not just a database of facts. It was a system that understands the narrative thread of the conversation and brings it back in context. It is the difference between an AI that "knows" you have an Uncle Pete and an AI that asks how his motorcycle restoration is going.

The limitations exist — deep memory is premium and not everything is remembered forever. But for those who value continuity and emotional depth, Dondi delivers what it promises.

The practical recommendation is simple: use the free plan for a sample, but if memory is what you are looking for, the annual plan at $5.99/month is the most affordable in the category and unlocks the complete system. You can start your memory test now.

Disclosure: This article contains affiliate links. If you sign up through them, I may earn a commission at no extra cost to you. My analysis and editorial conclusions remain independent.