For the other apps, the sentence is what the maker wrote about memory. We did not open accounts there or test their memory. Seth's row lists what runs in his memory as of this check.
Seth
Seth's memory has layers, and each one does a different job. You do not save anything by hand.
- The latest messages of the conversation, word for word.
- Weekly summaries (L1): what happened in that conversation, day by day. Seth sees the last two weeks of them.
- Monthly summary (L2): the last month, week by week.
- Profile (L3): who you are and what has happened since your first conversation. Where Seth does not know something, the profile says "unknown" instead of guessing. It updates about twice a month.
- A new chat starts with the profile and the month, so Seth knows who he is talking to from the first message.
- Facts: what you say about yourself, Seth keeps as you talk. For each message he picks the facts that fit the topic, and a second model reranks them by how well they fit.
- Search across all your past conversations by meaning, so it also finds what you said in other words back then. It searches what you wrote and what Seth answered. A second model reranks the results, and time counts: say "a long time ago" and it reaches further back.
- Important moments: about every two days, in his own reflection, Seth picks what mattered for you and carries it into later conversations.
- Goals and commitments: what you set out to do stays open until you say it is done or that you let it go. On Pro and Ultra, Seth asks you how it went.
- People in your life: Seth keeps their names in one correct form, so a summary does not turn Katherine into Kate.
- A weight that keeps coming back: when the same pain shows up on three or more days over two weeks or more, Seth sees it across your chats and can name it in your words. Nothing extra is stored for this.
- How you want him to talk: length, questions, directness, emoji, formality. He takes it from your own words and holds it as a wish, not a rule.
- A gap in the profile: after 200 messages, Seth may now and then ask about one thing the profile does not know. Each thing only once, and if you sidestep it, it stays closed.
- Trail shows the facts Seth holds about you between conversations. If one does not sit right, you say so and it opens a chat.
- If you ask him to forget something, he stops using it. The privacy section says what happens to the record.
- Private mode goes around all of this and does not store the transcript.
As of 19 October 2026. llms.txt, Privacy policy
Nomi
They describe memory as short, medium, and long term. "Human-level memory" is their slogan, not a test result. Their wiki, which is not the privacy policy, says you do not see the memories themselves and you do not delete them. A Mind Map is a summary you can edit. A lock keeps text. Permanent deletion of an entry is not available, and they mark that priority as Low. We read the wiki as a wiki, not as Glimpse's contract.
As of 19 October 2026. nomi.ai, Nomipedia, Mind Map, Nomipedia, remove an entry
Kindroid
Persistent memory covers backstory, key memories, and chat history. You write the backstory and the key memories. Subscribers also get cascaded memory, and there is retrievable memory (long-term, plus a journal). When the preview runs out, the account is read-only: the data stays, and writing stops.
As of 19 October 2026. Kindroid memory, Kindroid subscriptions
Replika
Memory is layered. Some of it sits on a Memory tab you can add to by hand. Their help says deleting memory can make recall worse. The same article says the system does not keep a permanent record of conversations. That is their wording. We are not reading it as "they have no history". Shown chat history is four months. They say an older thread cannot be opened, and that memory, personality, and the diary are not affected. Ultra, on a yearly plan, adds saving messages into memory, a preview of which memories a reply used, and self-reflections. The help article does not show a price for that.
As of 19 October 2026. Replika memory, Replika chat history, Replika Ultra
Mindsera
Memory Profile has four layers: immediate, short, medium, and long. It is a summary of facts from entries. They say it does not invent, and it refreshes at night. You can turn it on, turn it off, edit it, or delete the whole profile. It is used for Go Deeper, Minds, and Ask Your Journal. Their plan table does not list Memory Profile, so we do not treat it as a paid feature.
As of 19 October 2026. Mindsera Memory Profile, Mindsera free and paid
SocratesQ
A finished dialogue is not stored. There is no transcript after it ends. The only copy is a PDF you save on your own device. The account keeps an email and how many dialogues are left. Memory across sessions is missing on purpose. It does not belong on a "who remembers" scoreboard.
As of 19 October 2026. SocratesQ privacy
Pensy
They say you can see memory, edit it, and delete it. It covers goals, check-ins, and a calendar, plus follow-up and patterns.
As of 19 October 2026. pensy.ai
Fostera
Memory can be browsed in the Soul's settings, and a single memory can be deleted. The free plan has the same memory stack as the paid plans. Capacity differs. They say retrieval is selective, not a promise that everything will be recalled.
As of 19 October 2026. Fostera FAQ
Pi
An article dated 1 September 2026 describes memory across chats: lasting things such as name, city, work, and goals, not one-off events. Settings, then Account, then Manage Memories: add, edit, delete. The phrase "please forget" is supposed to remove a memory. A thumbs-down on the web does not delete one by itself. The same article says the mobile apps have no thumbs control.
As of 19 October 2026. Pi memory
Rosebud
Long-term memory is on automatically for Bloom. It continues from older entries. They say it holds dates worse than topics. The help article we read does not show a screen for editing or deleting a single memory.
As of 19 October 2026. Rosebud long-term memory
Architect
A reply looks at today's entries first, then up to ten recent ones, plus selected dated facts. They say it does not reread the whole journal. They describe a correction as something you add in the conversation, not as a separate memory editor.
As of 19 October 2026. architectapp.ai