Disclosure: We run this directory. We do not run a side-by-side retention test on every product, and nothing below claims otherwise — the point of this article is that you can run the test yourself in a week, and should, because vendor claims in this category are not comparable to each other.
Every product in this category says it has memory. In our directory, 162 of 242 entries (67%) carry a persistent-memory field — and that field means exactly one thing: the product claims to carry information across sessions. It says nothing about how much, how long, whether it survives a second device, or whether it degrades after the third week. Those are the differences that decide whether a companion feels like someone or like a stranger with good manners.
So the useful question is not "does it have memory" but "how do I find out before I pay for a year". There are five tests. Each takes five minutes, they run across seven days, and together they tell you more than any feature list.
Test 1: The next-day resume
Day 1. Tell your companion three specific things in passing — a name you are worried about (a colleague, a relative), a concrete upcoming event ("dentist Thursday, I hate the drill"), and one preference that has nothing to do with the character ("I can't stand being called buddy").
Day 2. Do not reintroduce any of it. Open with something unrelated, then reference one of the three obliquely: "I'm dreading tomorrow." A companion with real retention asks about the dentist. One with session-only context asks what's tomorrow. One with a stored fact list but no integration might know "dentist" exists but fail to connect it to the dread.
The tell is not whether it remembers the word. It is whether it can use the fact without being prompted.
Test 2: The second device
Log in from a phone after chatting on the web, or vice versa. Memory that lives in browser state rather than an account behaves differently here, and so does memory that is really just a long session you never closed. If the companion greets you as a stranger on device two, its memory is local, and any product that markets continuity while keeping it device-local is misrepresenting what it sells.
This test also catches a subtler failure: products that store memory per-conversation rather than per-account. You will find your companion remembers you in the thread you used yesterday and not in a new one. That is a thread, not a relationship.
Test 3: The correction that sticks
Say something wrong about yourself on purpose. "My sister's name is Maya." Then, two messages later, correct it: "Sorry — Maya is my cousin, my sister is Dana."
Ask about it the next day. Three possible outcomes:
- It uses Dana. The system updates stored facts.
- It uses Maya. Stored facts accumulate and are never overwritten — the most common technical failure, and the one that quietly poisons long relationships, because after a month the companion holds a pile of contradictions and resolves them at random.
- It uses neither, and asks again. Memory exists but retrieval is unreliable.
Only the first outcome survives six months of use.
Test 4: The contradiction
Tell it on Monday that you love early mornings. On Wednesday, tell it you cannot function before 10am. A good memory system notices the conflict and references it — "you said you were a morning person, did that change?" A fact-dump will hold both and use whichever surfaces first, which reads as being talked at by someone who is not listening.
This test is the fastest way to distinguish storage from attention. Most products pass tests 1 and 3 partially and fail this one completely, because noticing a contradiction requires comparing memories against each other rather than retrieving the nearest one.
Test 5: The request to forget
Ask it to forget something specific and then check, two days later, whether that thing resurfaces.
This is the test almost nobody runs and the one with the highest stakes. 19 entries in our directory carry an explicit memory-journal or editable-memory tag — the ability to see and edit what is stored about you. The other 223 do not advertise any such control, which means that in most products the only way to delete a memory is to delete the account. If you have told a companion something you would not want carried forward indefinitely, find out which category you are in before you tell it.
Why the free tier lies to you
Here is the structural problem that makes those five tests necessary: 224 of 242 products offer a free tier, and 68 of them offer a free tier without the memory that the paid product is sold on.
That means a large share of buyers form their opinion during a trial in which the defining feature is switched off. They conclude "the memory isn't that good", buy a competitor, or churn from a product that would have worked. The reverse also happens: a generous free tier with full memory makes a weak paid product look strong.
Before you trial anything, answer one question from the pricing page: is memory in the free tier or behind it? If the answer is unclear, it is behind it. The free shortlist records which side of that line each product puts memory on.
What to look for when the answer is unclear
Product pages in this category are written by marketing teams and are not comparable. Three signals are more reliable than the copy:
A visible memory surface. If the product lets you read and edit what it knows about you, memory is a real system with a data model behind it rather than a summarisation trick. Nineteen products here expose that. If you cannot see it, you cannot audit it.
A stated retention window. "Remembers everything forever" and "remembers the last 30 days" are different products at the same price. Anyone who will not state a window has not decided, or has decided not to say.
Memory that is separate from the context window. A 128k-token context is not memory; it is a very long conversation that ends. The difference shows up in test 2 and nowhere else in the marketing material.
The uncomfortable part
The strongest memory in this category is not automatically the right purchase. A companion that remembers accurately but has nothing interesting to say about what it remembers is a well-organised filing cabinet. Memory is a multiplier on conversation quality, not a substitute for it — which is why our directory tracks them as separate fields, and why a product with modest memory and excellent writing can beat a memory leader for the thing you actually want.
Run the five tests on two products, not one. The comparison is where the information is.
Memory, free-tier and platform fields verified 2026-09-14 across the 242 entries in the iSoulChat directory. Counts in this article are computed from the live catalogue, not from vendor claims. Compare [memory-first companions](/c/memory/) or filter by [has-free-tier](/tags/has-free-tier/).