AI Companions With Memory: Nomi, Kindroid & More Compared

“Remembers you” is one of the most important—and most ambiguous—claims in AI companionship. A chatbot may recall the last few messages fluently and still forget a detail from yesterday. Another may preserve an explicit character backstory but fail to mention it naturally. Before paying for “long-term memory,” find out which promise matters to you.

Five different things called memory

1. Immediate context: recent text provided to the model in the current conversation. It can fall out of scope as conversations grow.
2. Pinned facts and backstory: details you deliberately enter into fields meant to persist, such as names, preferences or roleplay canon.
3. Conversation summaries: compact versions of older discussions, useful for general continuity but not perfect word-for-word transcripts.
4. Retrievable memories: information stored over time and surfaced when the system judges it relevant. Storage is not the same as successful retrieval.
5. Editable learned context: a visible representation of how the app summarizes changing relationships and facts, when the provider offers one.

What different services actually document

Nomi describes short-, medium- and long-term memory, alongside Backstory+ and an evolving Identity Core. The provider explicitly notes that a memory can exist without being retrieved at the right moment. This is a more useful statement than a blanket claim of perfect recall.

Kindroid documents persistent fields, cascaded medium-term memory and long-term retrieval with journals. Its paid tiers differ in capacity, while its guidance notes that memory recall remains imperfect. Character.AI describes Story Memory and pinned facts, which are useful controls but should not be mistaken for flawless episodic memory.

The buyer question to ask

Do you want the companion to know a fixed profile of you, continue an ongoing fictional story, or remember seemingly incidental events from weeks ago? Those three goals require different evidence. If roleplay consistency matters, editable backstory may be more useful than a vague “infinite memory” claim. If ordinary conversation continuity matters, repeated recall tests are more relevant.

A useful memory test

Choose three fictional facts: a durable preference, a one-off event and a correction to an earlier assumption. Spread them across separate sessions and return to each with a neutral prompt. Do not give the exact answer away in your question. Log whether the system recalls correctly, asks for clarification or confidently invents an answer.

Follow the complete seven-day memory protocol and understand why context windows are not permanent memory. Then inspect the product research for Nomi and Kindroid.

A caution about simulated familiarity

Conversational models can sound confident and affectionate while getting basic facts wrong. Feeling recognized is not a measure of factual accuracy. Treat memory as a technical feature to evaluate separately from the emotional tone of the response.

Sources and evidence trail

These are primary-source documents reviewed for this research article. They describe what providers say their products do; they are not independent performance tests.

• Nomi: how memory works
• Kindroid: five memory systems
• Character.AI: 2026 memory update

Source review: 19 October 2026. Features, pricing and terms may change. Recheck before subscribing.