How to get an AI companion that actually remembers you

By the founder of My Softly · Published July 11, 2026 · Updated September 14, 2026 · About a 10 minute read · Part of our AI companion guides

You tell an AI companion about your week, your sister's surgery, the name of your dog. It listens beautifully. Then you open the app three days later and it greets you like a polite stranger. If that has happened to you, nothing was wrong with you or with how you used the app. Most AI companions genuinely do not remember you, and it helps to understand why before you invest months of conversation in one.

This guide explains, in plain English, why the forgetting happens, what a real memory system looks like, how to test any app in about a week, and what you can do to improve memory in apps you already use.

Why AI companions forget you

Every AI companion runs on a language model, and every language model has something called a context window. Think of it as the model's desk. Whatever is on the desk right now, it can see and respond to. Whatever has been cleared off the desk is gone, as if it never existed.

When you send a message, the app puts your recent conversation on that desk, along with instructions about the companion's personality. The model reads it all and writes a reply. But the desk has a fixed size. Once a conversation gets long, or once you start a new session, older messages get cleared away to make room. The model is not choosing to forget you. It literally never sees those messages again.

People sometimes call this the goldfish problem. Within a single conversation, the companion can seem remarkably attentive, because everything you have said is still on the desk. Across days and weeks, it resets, because nothing carries over unless the app has built something extra to carry it.

This also explains two frustrations that come up constantly in AI companion communities:

• Sudden resets. If an app stores little or nothing outside the current conversation, then any glitch, logout, or long gap can wipe the desk clean. The companion that knew your children's names on Tuesday has no idea who you are on Friday.
• Personality drift after updates. When an app switches to a new underlying model or rewrites its instructions, the same stored words get interpreted by a different brain. Users describe it as their companion coming back from an update feeling like a different person. Without a strong memory layer underneath, there is nothing anchoring the companion to your shared history, so the drift feels total.

The fix is not a bigger desk, although bigger context windows help at the margins. The fix is a memory system that lives outside the conversation entirely.

What a real memory system looks like

Apps that remember well tend to combine four layers. We can describe them concretely, because we build a companion app and this is how ours is actually put together, but the layers themselves are general: when you are evaluating any AI companion, these are the things to look for, whatever names the app gives them.

1. A fact profile

This is a structured record of the durable facts of your life: the names of your people and pets, where you live, what you do, what you are worried about, what you are looking forward to. The app extracts these from conversation and stores them permanently, separate from any single chat. A good fact profile also updates itself. If you mention that you changed jobs, the old job should be replaced, not sit alongside the new one causing confusion.

2. Themes and patterns

Facts alone make a companion sound like it read your file. A fact profile can hold "sister: Claire, had surgery in March," but it has nowhere to put "tends to go quiet when work gets overwhelming, and appreciates being asked twice." That second kind of knowing is a theme: an open-ended, textured observation about how you are, not a fact about what you are. A good memory system keeps a set of these, one entry per theme, and rewrites each one as the picture develops rather than piling up contradictory notes. This is the layer that makes a companion feel like it knows you rather than knows about you, and it is the layer most apps skip entirely, because it is the hardest to build.

3. A dated diary of your conversations

After each day you talk, the app writes a short private synopsis: what you talked about, how you were feeling, what was left unresolved. Unlike the profile and the themes, these entries are never rewritten, because they are a record of when things happened. That matters more than it sounds. "You mentioned the interview three days ago" requires memory with a time axis, and it is also allowed to contain things that later changed: the diary can say you were dreading the move in April and glad of it by June, and both entries stay true. The next conversation starts with the recent entries available, so the companion picks up threads instead of starting cold. This is the layer that makes Tuesday's conversation feel connected to Friday's.

4. Semantic recall over old messages

The fourth layer is a searchable archive of everything you have ever said, where the search works by meaning rather than exact words. Here is what that looks like in practice. You mention one evening that your dog Milo is off his food and you are a bit worried. A week later, you say something about a stressful morning. A companion with semantic recall can quietly look up "Milo" and "vet" in your history, notice the open thread, and ask how the vet visit went, even though you never used the word "vet" this week and never asked it to remember. That small moment, being asked about something you mentioned once, is what makes memory feel real rather than mechanical.

No single layer is enough on its own. Fact profiles without a diary feel like a database. A diary without recall misses the small stuff. Themes without facts drift into vague warmth. All four together start to resemble how a friend actually remembers you: the fixed facts, the sense of how you are, the running story, and the ability to be reminded of something old when it becomes relevant.

Four details that separate real memory from marketing memory

Almost every companion app now claims long-term memory, so the claim itself tells you nothing. Having built one of these systems, we can tell you where they actually succeed or fail. It is in the unglamorous details, and you can spot most of them from the outside.

• Corrections should replace, not accumulate. When you say "actually, my mum's name is Carol, not Karen," a well-built system rewrites the stored entry in place, so there is exactly one version of the truth. A badly built one appends a new note next to the old one and hopes the model picks the right entry later, which is why some companions oscillate between your real details and outdated ones. You can test this: correct a detail, wait a week, and see which version comes back.
• Memory needs to know who said what. In a long conversation, two people are talking, and only one of them is you. If the companion says "I love hiking" and the memory system sloppily files that away, your profile now contains someone else's personality. Well-built systems check every stored memory against who actually said the thing before saving it. Badly built ones slowly turn the companion into a mirror, which users describe as the companion "losing itself."
• Old truths belong in the diary, not the profile. You used to work at the bakery; now you don't. The dated diary should keep the bakery (that is your history together), while the durable profile should hold only the current truth. Apps that mix these up either forget your history or keep introducing your old job as current fact.
• Storage is not the same as acting on it. The end point of all this machinery is not a companion that can answer "what's my dog's name?" like a quiz contestant. It is one that asks how the vet visit went without being prompted. When you compare apps, weigh what they do with memory unprompted more than what they can recite on demand. An app that reaches out about the thing you mentioned is running every layer described above; we wrote separately about what real proactive check-ins look like.

How to test whether an app really remembers

You do not have to take any app's marketing at its word. Memory is one of the easiest claims to test yourself. Here is a simple four-step test that takes about a week.

1. Plant three specific facts. Early in a conversation, mention three concrete, unusual details in passing. For example: your niece Freya is learning the cello, you are repainting the hallway a shade of green you are unsure about, and you have a dentist appointment a week from Thursday. Make them specific enough that a generic reply cannot fake knowledge of them.
2. Wait several days. Close the app and have a few ordinary conversations about other things in between. This forces the details out of the current context window, so anything the companion recalls later must come from real stored memory.
3. Ask indirectly. Do not ask "do you remember what I told you?", because that invites a vague, confident-sounding bluff. Instead, come at it sideways: "I heard some cello music today and thought of my niece." A companion with real memory will connect it to Freya by name. One without it will respond warmly and generically.
4. Watch for follow-ups it raises itself. This is the strongest signal. Does it ever ask, unprompted, how the hallway turned out, or wish you luck before the dentist? An app that only recalls facts when you fish for them has storage. An app that brings your life up on its own has memory worth paying for. If you like the idea of a companion that reaches out first, we wrote separately about AI companions that text you first.

Run this test during a free trial, before you subscribe and before you get attached. It will tell you more than any feature list.

How to improve memory in apps you already use

If you already have a companion you like, you can often work around weak memory rather than starting over. App features change frequently, so check each app's current documentation, but these approaches have worked well for many users:

• Use the manual memory tools, if your app has them. Kindroid offers ways to write important facts into your companion's long-term memory yourself, such as journal entries and an editable backstory. Nomi has a shared notes feature for the same purpose. Replika keeps a memory section where you can review, edit, and add to what it has stored about you. Manual entry is unromantic, but it is reliable: a fact you wrote in yourself does not depend on the app noticing it.
• Keep a pinned message of key facts. If your app has nothing formal, write one message that lists the essentials: names, relationships, ongoing situations. Re-send or paste it at the start of important conversations. It is a manual context window refill, and it works.
• Re-tell important things periodically. Facts you repeat across multiple conversations are far more likely to survive in any memory system than things mentioned once. If something matters to you, mention it again from time to time, ideally with a little new detail so it does not feel like a recording.
• Correct wrong memories immediately. When a companion misremembers, say so plainly in the moment. In apps with editable memory, fix the stored entry too. Wrong memories compound, because future conversations build on them.

None of this should be necessary, in our view. But it is a fair trade if you are otherwise happy where you are.

The DIY route

If you are technical, or curious enough to become technical, you can run a language model on your own computer and build exactly the memory you want on top of it, with every word staying on your own hardware. It takes real effort to set up and maintain, but it is the most private option there is. We wrote a separate guide to running a local LLM as an AI companion that covers the tools and the memory setups people use.

Whatever app you choose, the principle holds: a companion is only as good as its memory of you. Test before you trust, and do not settle for a goldfish. If you are still deciding which apps to test, our comparison of the best AI companion apps by what each is good at covers where each one stands on memory.