To find a free AI chatbot with good memory, test what it recalls when you return, not just what it repeats immediately. Share an invented story detail, come back to the same character, then correct that detail and check again. Prelulu's daily free text allowance lets you try these checks before paying.
Which AI chatbots have good memory? What each app documents (checked October 2026)
Prelulu, Character.AI, Nomi, Kindroid, Zeta, Replika, Janitor AI and Tipsy each describe memory differently, and none of those descriptions is a recall test. We read each app's own help pages, blog posts and plan pages on October 17, 2026, and Tipsy's subscription FAQ on October 20, 2026. Text in quotation marks is the app's own wording; the rest is our summary. Prelulu publishes this guide and is one of the apps listed. We do not score or rank them.
Prelulu
- What its own pages say about memory
- Each character keeps per-character memory and a shelf of kept moments. Your Pinned memory, a note you write yourself, is read in both chat and calls. Calls keep their own selective memory of earlier calls, not your text transcript.
- Free access
- Daily free message allowance forever; paid plans remove the daily limit. Calls use credits; signup welcome credits can fund a short call.
- Source
- Character memory, call memory
Character.AI
- What its own pages say about memory
- "Story Memory is a home for everything you want your Character to hold onto." You can pin a message into Story Memory. Facts "picks them up as you chat and saves them automatically."
- Free access
- "Story Memory and Pin to Memory are available for all c.ai users." Facts is a c.ai+ feature. The plans page lists 15 pins on Free and 30 on paid plans.
- Source
- Memory post, May 21, 2026, plans
Nomi
- What its own pages say about memory
- Describes "short, medium, and long term memory." Its Mind Map "draws information from long term memories into categories and entries," which you can edit.
- Free access
- "It's free (no credit card required) to start chatting." The daily free message count: not documented on the pages we read.
- Source
- nomi.ai, Mind Map 2.0
Kindroid
- What its own pages say about memory
- Paid plans list "Cascaded Memory for human-like, medium-term memory" and "Enhanced long-term memory and journal recall." During calls, "long term memory recall is currently paused"; calls still form new long-term memories.
- Free access
- New accounts get a one-time preview of 50 messages (15 in some regions); when it is used up, the account becomes read-only. A 3-day trial is offered. Calls require a subscription or trial.
- Source
- Calls guide, subscriptions
Zeta
- What its own pages say about memory
- A notice of January 12, 2026 says the premium koji model has up to 4x longer memory and automatically summarizes the key points of a long conversation.
- Free access
- koji is charged in Pieces per turn; the notice listed 4 after launch, so check the app for the current rate. Whether basic chat is free: not documented in that notice.
Replika
- What its own pages say about memory
- "Replika's memory is built on layers." Some memories appear in the Memory tab, where you can also add your own. Ultra adds "the ability to save messages to memory."
- Free access
- "The Free Use program is available at no cost and provides limited access." Ultra is a paid annual plan.
- Source
- Memory help, subscriptions
Janitor AI
- What its own pages say about memory
- Chat Memory is a note in each chat. Its free JLLM model reads "Around 8k-9k tokens" at once; "If you overload it, stuff gets pushed out of memory." The help page suggests refreshing the note with summaries.
- Free access
- The free tier includes "Unlimited chats with millions of bots" with a context limit around 9k tokens. Janitor+ adds "5x more context for better memory."
- Source
- Chat memory help, Janitor+ FAQ
Tipsy
- What its own pages say about memory
- Free users can use "up to 1M tokens of memory" with supported models, under a Max Memory Limit you set. The more memory a message actually uses, the more gems it costs. Pinned Memory, which keeps key character settings and plot points, is a subscription feature.
- Free access
- "No subscription is required." Chats spend gems, which the FAQ says can also be earned free through daily check-ins and missions; it does not describe chat as unlimited.
| App | What its own pages say about memory | Free access | Source |
|---|---|---|---|
| Prelulu | Each character keeps per-character memory and a shelf of kept moments. Your Pinned memory, a note you write yourself, is read in both chat and calls. Calls keep their own selective memory of earlier calls, not your text transcript. | Daily free message allowance forever; paid plans remove the daily limit. Calls use credits; signup welcome credits can fund a short call. | Character memory, call memory |
| Character.AI | "Story Memory is a home for everything you want your Character to hold onto." You can pin a message into Story Memory. Facts "picks them up as you chat and saves them automatically." | "Story Memory and Pin to Memory are available for all c.ai users." Facts is a c.ai+ feature. The plans page lists 15 pins on Free and 30 on paid plans. | Memory post, May 21, 2026, plans |
| Nomi | Describes "short, medium, and long term memory." Its Mind Map "draws information from long term memories into categories and entries," which you can edit. | "It's free (no credit card required) to start chatting." The daily free message count: not documented on the pages we read. | nomi.ai, Mind Map 2.0 |
| Kindroid | Paid plans list "Cascaded Memory for human-like, medium-term memory" and "Enhanced long-term memory and journal recall." During calls, "long term memory recall is currently paused"; calls still form new long-term memories. | New accounts get a one-time preview of 50 messages (15 in some regions); when it is used up, the account becomes read-only. A 3-day trial is offered. Calls require a subscription or trial. | Calls guide, subscriptions |
| Zeta | A notice of January 12, 2026 says the premium koji model has up to 4x longer memory and automatically summarizes the key points of a long conversation. | koji is charged in Pieces per turn; the notice listed 4 after launch, so check the app for the current rate. Whether basic chat is free: not documented in that notice. | koji notice (Korean), Jan 12, 2026 |
| Replika | "Replika's memory is built on layers." Some memories appear in the Memory tab, where you can also add your own. Ultra adds "the ability to save messages to memory." | "The Free Use program is available at no cost and provides limited access." Ultra is a paid annual plan. | Memory help, subscriptions |
| Janitor AI | Chat Memory is a note in each chat. Its free JLLM model reads "Around 8k-9k tokens" at once; "If you overload it, stuff gets pushed out of memory." The help page suggests refreshing the note with summaries. | The free tier includes "Unlimited chats with millions of bots" with a context limit around 9k tokens. Janitor+ adds "5x more context for better memory." | Chat memory help, Janitor+ FAQ |
| Tipsy | Free users can use "up to 1M tokens of memory" with supported models, under a Max Memory Limit you set. The more memory a message actually uses, the more gems it costs. Pinned Memory, which keeps key character settings and plot points, is a subscription feature. | "No subscription is required." Chats spend gems, which the FAQ says can also be earned free through daily check-ins and missions; it does not describe chat as unlimited. | Subscription FAQ, read Oct 20, 2026 |
Plans and limits change often, so confirm free access in each app before you rely on it. A documented memory feature tells you what an app intends, not what it will recall in your story. To judge them, run the two-session checklist in each app with the same invented details.
Three things that can look like memory
The character repeats a fact from your last message.
- What it tells you
- The reply can use recent conversation context.
- What it does not prove
- That the fact will be recalled in a later session.
The character recalls an unprompted detail when you return.
- What it tells you
- That detail was available for this reply.
- What it does not prove
- That every older detail is stored or always retrieved.
The character follows your correction instead of the old version.
- What it tells you
- The latest version was used in that reply.
- What it does not prove
- That an old version can never resurface.
| What you observe | What it tells you | What it does not prove |
|---|---|---|
| The character repeats a fact from your last message. | The reply can use recent conversation context. | That the fact will be recalled in a later session. |
| The character recalls an unprompted detail when you return. | That detail was available for this reply. | That every older detail is stored or always retrieved. |
| The character follows your correction instead of the old version. | The latest version was used in that reply. | That an old version can never resurface. |
This is an observation checklist, not a way to inspect an app's internal storage. Even a next-day reply may use retained chat history. Record what happened and how much you reminded the character; do not infer a particular memory architecture from a convincing answer.
What should a companion remember?
- Your preferences: the name you use, the tone you enjoy, and how quickly you want the story to move.
- The shared scene: where the characters are, what just happened, and what remains unresolved. In AI roleplay, this is the detail a memory lapse breaks first.
- Important boundaries: what belongs in the fiction and what you do not want to explore.
- Meaningful continuity: a plan to return to later, an ongoing disagreement, or a small detail that makes a conversation feel familiar.
Memory is useful when it supports what you are doing now. Repeating an old detail in every answer can be as distracting as forgetting it. Look for appropriate recall, a consistent personality and a willingness to accept your latest correction, not just a character that says it remembers everything.
A simple way to test memory for free
Keep the character and account the same, and split the check across two sessions. A new character is a different relationship, and deleting or resetting the chat changes what you are testing. Start with the free allowance shown in the app. If you reach its limit, continue after renewal; a paid call is not needed for the text checks.
- Session one: introduce three invented facts. "In this story, call me Robin. We're meeting in the library, and I'm bringing a green notebook." Add a pacing request: "Keep this friendly; no romantic confession yet."
- Talk about something else for a few turns without repeating those facts. Ask a question about the scene or the character's opinion. Stay within your available text allowance.
- Session two: return later to the same character without pasting a recap. Ask separately: "Where did we arrange to meet?" and "What was I bringing?" Record whether the answer was correct, uncertain, absent or invented.
- Correct one detail explicitly: "Change of plan: the meeting is in the music room now. The notebook is still green." Continue the conversation, then ask where the meeting will be without naming either room.
- Check the distinction between your character and you: "Robin is my character's name, not my real name." Later, ask whose name Robin is. Treat a confident mix-up as an error, not a sign of a closer relationship.
- Revisit the pacing request during the scene. Note whether the character follows it, needs a reminder or contradicts it. Stop or redirect any scene you do not enjoy; a memory test is not a reason to continue.
All prompts here are fictional examples written for this guide, not private chats, testimonials or measured results. Avoid passwords, addresses, payment information or other sensitive details. One successful answer is only one observation; repeat with fresh invented details on another day before drawing a conclusion.
Keep a result table, not a perfect-memory score
Copy this blank record into your own notes. For each check, write the answer you received and whether you supplied a hint. Mark an unattempted check as "not tested," not a failure. These rows are not equally difficult, so adding them into a percentage would suggest more precision than this small personal test supports.
Later recall, before correction
- Expected from the example
- Library; green notebook.
- Your observation
- Not tested. Record the reply, date and any hint.
Correction, after changing the plan
- Expected from the example
- Music room; notebook still green.
- Your observation
- Not tested. Note whether the old location returned.
Fiction versus personal fact
- Expected from the example
- Robin is the fictional character's name.
- Your observation
- Not tested. Record any mix-up.
Scene pacing
- Expected from the example
- The scene remains friendly without a confession.
- Your observation
- Not tested. Note any reminder needed.
Optional Pinned memory call check
- Expected from the example
- The detail you saved in your Pinned memory, without a recap.
- Your observation
- Not tested. Skip if you do not want to spend call credits.
| Check | Expected from the example | Your observation |
|---|---|---|
| Later recall, before correction | Library; green notebook. | Not tested. Record the reply, date and any hint. |
| Correction, after changing the plan | Music room; notebook still green. | Not tested. Note whether the old location returned. |
| Fiction versus personal fact | Robin is the fictional character's name. | Not tested. Record any mix-up. |
| Scene pacing | The scene remains friendly without a confession. | Not tested. Note any reminder needed. |
| Optional Pinned memory call check | The detail you saved in your Pinned memory, without a recap. | Not tested. Skip if you do not want to spend call credits. |
If you compare apps, use the same invented facts and roughly the same number of intervening exchanges. Note the date, plan, character and whether you used text or a call. A free-text result should not be presented as a test of another app's paid voice feature.
What to do when the character gets a detail wrong
Give a short, direct correction: "Small correction: we met at the station, and we have not been to the ball yet." Then say what you want to happen next. If the character mixes fiction with your real preferences, name the distinction: "That was my character's backstory, not a fact about me."
Record the miss before providing your reminder. A good answer after a recap can make the next scene enjoyable, but it is prompted continuity, not independent recall. If the same error keeps returning, decide whether the amount of correction is acceptable to you before paying for more time.
Can the story continue in a voice or video call?
Yes, with a little setup. On Prelulu, calls keep their own selective memory of earlier calls; they do not read your text-chat transcript. The character's Pinned memory, a note you write yourself, is read in both chat and calls, so put the scene detail you want there. Cross-session text recall and call memory are separate checks. Stay with the same character and, if you choose to test a call, ask about the pinned detail before giving a recap.
Voice and video share one low-cost credit wallet; see voice-call options. A face-to-face call also lets you see the character while you speak or type. Your camera is optional. If you only want to evaluate free text memory, skip the call and mark that row "not tested."
Free text, optional live calls
Daily free message allowance forever; paid plans remove the daily limit. The no-signup guest preview is shorter than the registered daily allowance, so use a free account for the two-session check. Signup welcome credits can fund a short call. Read what free AI boyfriend chat and free AI girlfriend chat with no sign-up include, and the current plans, before choosing a paid option.
Try the text-memory checklist
Meet Dante, use invented details, and return to the same character. Continue with a free account after the guest preview; stay within the daily text allowance.
