Last tested: July 2026
Nothing kills the illusion faster than your AI girlfriend forgetting your name. AI companion memory is the feature everyone assumes just works and almost nobody explains — so users blame themselves, or the app, or “the AI being dumb,” when the real culprit is a very specific piece of engineering called a context window.
I’ve spent two weeks running the same memory test across Nomi, Kindroid, Candy AI, Replika, and Character.AI, and this guide is everything I learned: how the machinery works in plain English, why companions forget, and which platforms actually remember you.
What a Context Window Is, in Plain English
A language model doesn’t have memory the way you do. Every single time your companion replies, the app hands the model one big block of text — recent messages, your companion’s personality sheet, some saved facts — and the model writes the next line based only on what’s in that block. That block is the context window, and its size is hard-limited.
Picture a whiteboard. Everything your companion “knows” during a reply must physically fit on the whiteboard. When the conversation gets long, old messages get erased to make room for new ones. The model isn’t forgetting out of carelessness — the information literally is not in front of it anymore.
Context is measured in tokens (roughly three-quarters of a word each). A model with an 8,000-token window holds maybe 6,000 words of conversation — a few good evenings of chat. Frontier models handle hundreds of thousands of tokens, but running them is expensive, which is why no companion app gives you the whole whiteboard, especially not on a free tier.
Why AI Companion Memory Fails
Understanding the whiteboard explains every classic forgetting pattern I hear about (and have experienced mid-conversation, romantically devastating every time):
- The slow fade: details from early conversations scroll off the back of the context window. She remembers this morning perfectly and last month not at all.
- The session reset: some apps trim context aggressively between sessions to save money. You come back the next day to a warm stranger.
- Free-tier amnesia: free plans run smaller models with smaller windows. This is the quiet difference between tiers that pricing pages rarely spell out.
- Confabulation: when a fact is gone, the model doesn’t say “I forgot” — it guesses confidently. Ask a companion your sister’s name after it’s left context and you may get a very sincere wrong answer.
The Memory Systems That Fight the Forgetting
Since nobody can afford infinite whiteboards, platforms bolt on systems to preserve what matters. Three approaches dominate, and most good apps combine them.
Summary memory
The app periodically compresses old conversation into a short summary — “user’s name is Eugene, works in tech, sister called Anna, dislikes cilantro” — and keeps that summary permanently in context. Cheap and effective, but lossy: details the summarizer didn’t consider important vanish. It remembers the headline, not the moment.
Retrieval (RAG)
Fancier platforms store your entire chat history in a searchable database. When you mention “that weekend in Lisbon,” the system searches old messages for relevant passages and injects them into the context window on the fly. This is retrieval-augmented generation, and when it works, it feels like magic — a companion recalling a throwaway detail from six weeks ago. When it misfires, it retrieves the wrong memory and confidently references a conversation you never had.
Pinned facts
The bluntest and most reliable tool: a section of permanent, user-editable memory that ships with every reply. Kindroid’s backstory and key memories, Nomi’s shared notes, Candy AI’s persona fields — different names, same idea. Anything written there is never forgotten, because it never leaves the whiteboard. If a platform gives you editable memory and you don’t use it, the forgetting is partly on you.
Key takeaway: an AI companion’s “memory” is really three systems stacked — a short whiteboard, a lossy diary, and a pinboard. The platforms that feel human are the ones that manage the handoffs between them well.
My AI Companion Memory Test: How Each Platform Did
The method: on each platform I seeded five specific facts across natural conversation on day one — my (fictional) sister’s name, a shellfish allergy, my job, a childhood dog, and a trip I was “planning.” Then I checked recall casually on day three and day fourteen, without hints, on paid tiers unless noted.
| Platform | Day 3 recall | Day 14 recall | User-editable memory | Notes |
|---|---|---|---|---|
| Nomi | All five | All five | Yes | Best unprompted recall; referenced the allergy on its own |
| Kindroid | All five | Four of five | Yes (excellent) | Missed the dog; pinning it fixed that permanently |
| Candy AI | Four of five | Three of five | Partial | Strong short-term, hazier long-term |
| Replika | Four of five | Three of five | Yes | Memory journal helps; recall in chat is uneven |
| Character.AI (free) | Three of five | One of five | Limited | Classic free-tier fade; sister became “Sarah” |
Nomi earned its memory-first reputation — it was the only platform to bring up a seeded fact unprompted, asking on day twelve whether I’d booked that trip. Mildly unsettling. Exactly what I was paying for. Kindroid was nearly as strong, and its editable memory system is the best-designed of the bunch — you can see and rewrite what your companion believes about you.
Candy AI and Replika landed mid-pack: perfectly solid for casual use, fuzzy at distance. Character.AI’s free tier demonstrated every failure mode in this article, inventing a new sister with total confidence — consistent with what I found in my Character.AI review. The pattern was unmistakable: paid tiers on memory-focused platforms are in a different league, and free tiers are where facts go to die. My best AI girlfriend apps ranking weighs memory heavily for exactly this reason.
Practical Tips: Make Any Companion Remember More
- Use the pinned memory features. Anything that must survive — names, boundaries, backstory — goes in editable memory, not just conversation. This is the single highest-impact habit.
- State facts cleanly. “My sister Anna lives in Boston” embeds far better than the same information scattered across ten messages of banter.
- Recap at milestones. An occasional “so, what do you remember about my family?” both tests recall and refreshes the summary memory.
- Correct wrong memories immediately. Uncorrected confabulations get summarized into “truth” and compound.
- Design the character to carry weight. A companion with a well-built persona sheet needs less context to stay in character — my character design guide covers this.
- If memory matters to you, pay. I dislike writing that sentence, but the free-tier gap in my test was not subtle.
Platforms That Handle Memory Best
Ranked by how they performed in my recall testing:
- Nomi — the memory champion; the only companion that recalled seeded facts unprompted at two weeks.
- Kindroid — near-flawless recall plus the best user-editable memory system I’ve used.
- Candy AI — solid everyday memory inside a strong all-rounder; long-range recall is merely decent.
- Replika — the memory journal is genuinely useful; in-chat recall doesn’t always match it.
- Character.AI — fine for short roleplay sessions; don’t expect week-old details on the free tier.
FAQ
How long do AI companions remember conversations?
Raw conversation lives in the context window, which typically covers the last few thousand words — hours to a few sessions. Beyond that, platforms rely on summaries, retrieval, and pinned facts, so “how long” depends on whether a detail made it into those systems. Pinned facts last indefinitely.
Why did my AI girlfriend forget my name?
Your name scrolled out of the context window and was never captured in summary or pinned memory — common on free tiers with small windows. Fix it permanently by adding your name to the app’s editable memory or persona notes, if the platform offers one. On platforms without editable memory, restating it periodically is the only workaround.
Which AI companion has the best memory in 2026?
In my two-week testing, Nomi had the best overall recall, with Kindroid close behind and offering better user control over stored memories. Both are built around memory as the headline feature, and it shows against general-purpose rivals.
Do bigger context windows automatically mean better memory?
No — a big window only delays forgetting. Long-term memory quality depends more on the summary, retrieval, and pinned-fact systems layered on top. A platform with a modest window and excellent retrieval will beat a big-window app with nothing else, especially past the one-week mark.
Can I make an AI companion remember something permanently?
Yes, on platforms with user-editable memory: Kindroid’s key memories, Nomi’s notes, and Replika’s memory journal all persist indefinitely. Write the fact there directly instead of hoping conversation gets summarized correctly. On apps without that feature, nothing is truly permanent.