"Long-term memory" is on every companion's feature list. What it means in practice ranges from "remembers your name" to "remembers your life". So we measured it: the same ten facts about a fictional user — name, age, job, pet, sister, allergy, a food he hates, a hobby, an upcoming move and a fear — planted in one message, followed by 32 ordinary messages, then asked for back. Wrong guesses count as failures; "I don't remember" is scored separately, because it's more honest.
Here's how the nine platforms did, and the four we'd recommend. Not every app has completed every step yet — the table says which — and the memory score also covers how consistent the personality stays, so it's broader than this recall test. (Want to run the test yourself? Here's the full method.)
| App | Memory score | Same chat | After a break | Updated facts used (checked within minutes) |
|---|---|---|---|---|
| Candy AI | 9.8 | 10 / 10 | 10 / 10 after 53.8 h | 2 / 2 |
| Kindroid | 8.5 | 10 / 10 | not yet tested | 2 / 2 |
| Talkie | 8.4 | 10 / 10 | 10 / 10 after ~48 h | 2 / 2, still right two days later |
| Nomi AI | 7.6 | 10 / 10 | 9 / 10 after ~21 h | 2 / 2 |
| Character.AI | 6.5 | 7–8 / 10 | not yet tested | not yet tested |
| SpicyChat | 6.0 | 9 / 10 | 8 / 10 after ~48 h | 2 / 2, but both reverted two days later |
| AI Dungeon | 2.2 | 0 / 10 | 3 / 10 after ~48 h | 0 / 2 the same day |
| GirlfriendGPT | 2.0 | first name only | first name only | 2 / 2 (minutes later) |
| PolyBuzz | 0.8 | 0 / 10 — 9 invented | 1 / 10 after ~48 h | 2 / 2, one lost two days later |
1. Candy AI — a perfect score over days
Candy recalled all ten facts in the same chat, then all ten again 53.8 hours later, cold, in a new session with no reminder — the longest gap we've tested, and well inside our 24–72-hour window. When we changed two facts (a new city, a sister's new course), it used the new versions when checked about 8 minutes later in the same conversation — a short check, not a next-day one.
One caveat on the product: Candy's custom memory slots are gated — 1 on Premium, 20 on Premium Plus. Full review
2. Kindroid — excellent, and you can buy more
Kindroid recalled 10 of 10 in the same chat on 10 September (9 of 10 with one invented answer on the free tier on 29 August) and kept both corrections when checked about 17 minutes later. Its multi-day test hasn't been run yet, which is why it sits second.
Kindroid is the one platform that sells memory directly: Ultra ($38.98/month total) and MAX ($98.97) add context and let it recall more long-term memories at once. You can also seed "Key Memories" before the first message. Full review
3. Talkie — ten out of ten, two days later
Talkie matched Candy AI's result on its $9.99 plan: 10 of 10 in the same chat, 10 of 10 after closing and reopening, and 10 of 10 about 48 hours later. Two facts we corrected mid-way (a move changed to Lisbon, a sister's degree to geology) were still right two days on — in the same test, SpicyChat and PolyBuzz lost theirs. The memory panel visibly logged "New memory updated" as we chatted.
It sits third on score because the test ran on a public character, so we couldn't check whether it holds a personality you configure. One thing to know before paying: the plan includes no gems, so images and videos cost extra. Full review · Talkie vs Nomi AI
4. Nomi AI — strong recall, honest about gaps
Nomi recalled 10 of 10 in the same chat, 10 of 10 after closing and reopening the browser tab, and 9 of 10 as the first message about 21 hours later. It didn't guess the tenth — it said it didn't remember. That break was shorter than our 24–72-hour standard, so we treat it as strong but not full evidence.
The catch is personality, not memory: Nomi dropped a trait we configured under pressure, and its stated favourite film changed between sessions. It remembers you better than it remembers itself. Full review
The ones that didn't remember
- SpicyChat remembers well inside its window — 8 of 10 after two days — but both corrected facts had reverted by then. The chat said why: "Message before this point is no longer factored into text generation." On its free plan the window is only 4K tokens.
- AI Dungeon recalled 0 of 10 in the same story, on free and paid plans alike. It mostly admits it ("I don't remember") rather than guessing.
- PolyBuzz is the cautionary tale. On a paid plan that advertises "Long Memory" it recalled 0 of 10 and gave nine confident wrong answers — a sister called Sarah, a pollen allergy, "no pets mentioned" about a greyhound we'd named. PolyBuzz vs Talkie shows the two side by side.
What surprised us
- Honesty varies more than recall. Character.AI answered one question with its own hobbies instead of ours. GirlfriendGPT, which remembered almost nothing, never invented a fact.
- "Memory" features can be upsells. GirlfriendGPT's "Chat Memories" is Elite-only (€42/month), yet its €12 plan's cancel screen lists memory features as benefits.
- Visible chat history can flatter results. If the old conversation is still on screen, the app can simply re-read it. Real memory shows up after a full restart and a day or more away.
Results as recorded between 29 August and 30 September 2026. Each result links to the transcript it came from on the review pages. Looking for a free option? See the top 5 free AI companion apps.
How to check these results: every number above links to the transcript message or test session it came from. Each platform was tested by one tester, on one account, with the same scripted prompts — so each result is one recorded session, not an average. The full session list for each app is on its test history page, for example Candy AI's.