Quick answer
- Short-term memory is the context window: the slice of recent conversation the model can read when it writes a reply. Anything outside it is invisible unless the app puts it back.
- Long-term memory means stored information about you or the conversation that the app keeps beyond one chat and brings back when relevant.
- Apps stack four tools: manual notes and pins, automatic summaries, searchable memory stores and lorebooks.
- Names differ: Mind Map (Nomi), cascaded and retrievable memory (Kindroid), Memory tab (Replika), Story Memory (Character.AI), Semantic Memory 2.0 (SpicyChat).
- No app remembers everything, and several say so.
"She remembers you" is the most common promise in AI companion marketing and the least precise. This explainer is part of our AI companion app guides. It describes how memory works in general, using vendor documentation, and then names the real tools in each app from help centers and plan pages checked on October 20, 2026. We did not create accounts or test recall. To see which apps do it best, use our ranking of AI companions with memory.
What Long-Term Memory Means in an AI Companion
In a chatbot, long-term memory means stored information about the user or the conversation that an app keeps beyond a single chat and brings back when relevant. As a general pattern (our summary of the vendor documentation), it is text kept in a database and re-inserted into the prompt, not the model permanently changing itself while you talk.
Vendors describe it in similar terms. OpenAI says ChatGPT "can remember relevant preferences and details from your chats", and adds that memory "does not retain every detail from every conversation". Character.AI calls Story Memory "a home for everything you want your Character to hold onto". Kindroid's update log says its retrieval models were upgraded so "the memories most relevant to the current conversation come up more often". The common thread is selection: the app decides which stored facts matter for your next message.
Short-Term Memory: The Context Window
A language model can only use the text placed in front of it. Anthropic describes the context window as "working memory" and Google compares it to short-term memory. Windows are measured in tokens, and apps ration them. SpicyChat's context is 4,096 tokens on its free plan, 8,192 on True Supporter and up to 32,768 on I'm All In; CrushOn.AI offers 8K, 16K or 24K by plan; Janitor AI's free model holds about 9K, with 5x more on janitor+.
A bigger window is not automatically better. Anthropic notes that accuracy and recall can degrade as the token count grows. And the window also holds the character definition, so a long persona leaves less room for chat. Our glossary entry on the context window defines the term, and the Replika detail is worth knowing: the app displays only the past four months of chat history, while memories and diary entries remain.
The Four Layers of Companion Memory
1. Manual memory you write
Notes, pins and backstory fields are sent with every reply. Character.AI allows 15 pinned messages free and 30 on its paid plans; CrushOn.AI pins 4,000 to 8,000 characters per chat by plan (about 800 per pin is recommended); Kindroid has a backstory and key memories; Janitor AI has a Chat Memory field you fill yourself. They are reliable because they are always present, and limited for the same reason: they use up the window.
2. Automatic summaries
When older chat no longer fits, some apps compress it. SpicyChat's Semantic Memory 2.0 (True Supporter and up) turns chats into memories automatically, CrushOn.AI's Auto-Summarization writes summaries into long-term memory with Resummarize on Luxe and above, and Character.AI compresses old context in the background. Muah AI's optional Memory Compaction (beta) writes an editable System Summary, and its guide warns that "a compact summary is not an archival backup". Summaries save space and lose detail.
3. Searchable memory stores
The most capable systems keep many entries and retrieve only the relevant ones for each message. Kindroid describes persistent memory (backstory, key memories, Learned Context), cascaded memory (a proprietary medium-term layer for subscribers) and retrievable memory, with unlimited long-term memories. Nomi's Mind Map 2.0 shows memory as an editable graph and table: 50 entries per category are eligible for recall, up to 10 can be marked high priority, and each entry holds up to 10,000 characters. Replika's Memory tab lists what it remembers, and you can add, edit or delete entries.
4. Lorebooks and keyphrase triggers
A lorebook stores facts that enter the prompt only when certain words appear. In the community character card specification, each entry has keys and content, and a book has a token budget and scan depth. Kindroid's journal entries work this way, with up to 8 keyphrases per entry and 500 entries per set. SpicyChat and Janitor AI support lorebooks, and Character.AI's Lorebook is in beta. See how AI roleplay works for how creators use them.
What Each App Calls Its Memory
| App | Names for the tools | What you can edit |
|---|---|---|
| Nomi | Mind Map 2.0, Identity Core, Backstory+ | Mind Map entries, Backstory+ sections |
| Kindroid | Persistent, cascaded and retrievable memory; Learned Context; journal | Backstory, key memories, journal entries |
| Replika | Memory tab, diary; Ultra saves memories | Add, edit, delete memories |
| Character.AI | Story Memory, pins, Facts (c.ai+), Memory Usage bar | Pins; Facts |
| SpicyChat | Memory Manager, Semantic Memory 2.0, lorebooks | Entries of 250 characters |
| CrushOn.AI | Pins, Auto-Summarization, Save Memories | Pins; up to 100 saved messages |
| Janitor AI | Chat Memory field, scripts and lorebooks | Chat Memory text |
| Muah AI | Memory Compaction (beta), System Summary | The summary |
| Kupid AI | Memory counted in messages: last 30, last 100, unlimited | None documented |
| Candy AI | Chat memory; 20 memory slots on Premium Plus | Reset or delete the chat |
| Others | Talkie says memory capacity is "technically limited"; DreamGF, GirlfriendGPT, Chai and Romantic AI document no memory system | Not documented |
Why Companions Still Forget
- The window filled up, and older text dropped out. Janitor AI's help center says the model forgets the oldest text when full.
- A summary lost the detail. Compression keeps the gist, not the exact line.
- Retrieval missed. Nomi's wiki says recall is not perfect, and an older Character.AI help article says characters "appear to forget things".
- The plan changed. Kindroid's cascaded memory is for subscribers, and CrushOn.AI deletes free-plan chat history after 7 days of inactivity.
- An update shifted behavior. One Trustpilot reviewer reported Kupid AI memory shrinking after an update (a single user report).
How to Help Your Companion Remember
This is general advice, not an app's instruction. Put stable facts such as your name, job and rules for the relationship in the editable fields, because those are always sent. Keep entries short and factual. Fix a wrong memory when you see it instead of arguing in chat. Save key story moments as pins or journal entries, and never treat a summary as a backup. Our creator guide on memory setup for AI characters shows how to lay these fields out, and free AI companions with memory lists what free plans keep.
Memory and Your Privacy
A memory store is a profile of you, and deleting it is not always simple. Muah AI's guide says deleting a message does not guarantee that a fact already in the summary disappears. Candy AI clears conversation and memory together when you reset or delete a chat, and Kupid AI says deleting a chat removes that companion's memory. Replika lets you delete single memories but not your whole message history without deleting the account. Our guides on whether apps store your chats and deleting an AI companion account cover the rest. For how memory fits with the model, voice and moderation, read how AI girlfriend apps work.
FAQ
What does long-term memory mean in an AI companion?
Stored information about you or the conversation that the app keeps beyond one chat and brings back when relevant. It can be notes you write, automatic summaries or a searchable store of facts, and it is separate from the context window.
Does my AI companion learn from me?
Memory features store text and re-insert it. Separately, some companies use chats to improve their models: Nomi says Nomis "in aggregate use information from chats to learn", with names removed, and Character.AI retains some data to help train its models.
What is a context window?
The amount of text, measured in tokens, that a model can consider when writing a reply. SpicyChat offers 4K to 32K tokens by plan and CrushOn.AI 8K to 24K.
Can I edit what my AI companion remembers?
On Nomi (Mind Map), Replika (Memory tab), Kindroid (backstory, key memories, journal), SpicyChat (Memory Manager, paid), CrushOn.AI and Character.AI (pins) and Janitor AI (Chat Memory). Muah AI lets you edit its summary.
Do free plans include long-term memory?
Partly. Nomi, Character.AI, Janitor AI and CrushOn.AI include memory tools on free plans, and SpicyChat's free plan has a 4K window, usually smaller than paid ones. Kindroid has no free plan since October 1, 2026.
Sources
Official pages and store listings we read for this page (last check October 20, 2026).
- platform.claude.com: Anthropic definition of context window
- ai.google.dev: Google definition of context window (short-term memory analogy)
- help.openai.com: OpenAI definition of Memory in ChatGPT
- blog.character.ai: Memory features 2026-05-21
- kindroid.ai: Help center: memory
- kindroid.ai: Official update log (Oct 1 free-tier change, V1 retirement)
- wiki.nomi.ai
- help.replika.com
- docs.spicychat.ai: Context and reply limits by tier
- docs.spicychat.ai
- aiwiki.crushon.ai: Official wiki: auto-summarization and memory
- crushon.ai: Pricing, Monthly and Annually tabs (EU IP, logged out)
- help.janitorai.com: Memory / context window
- muah.ai: Official user guide, reviewed October 3, 2026
- everai.zendesk.com
- kupid.ai: Pricing: Weekly, Yearly, Lifetime tabs clicked, Pro and Ultra cards selected
- talkie-ai.com: FAQ (NSFW, refunds, memory, model)
- github.com: Character Card V2 specification