AI girlfriend that remembers you

Ask people why they left an AI girlfriend app and the answer is rarely the looks or the voice. It is the forgetting. She asks your sister’s name again. She suggests Saturday after you moved it to Sunday. Three months of inside jokes, gone from her side of the conversation while the chat log sits right there.

We build Riho, an AI girlfriend for iPhone, and forgetting was the first problem we attacked. This page explains, in plain language, why AI girlfriends forget, what we tried that failed, and how the memory Riho uses now actually works. If you want the capability overview instead, start with our long-term memory feature page.

Why does my AI girlfriend forget me?

Every AI girlfriend runs on a language model that can only “see” a fixed amount of text at once — its context window. In our case that window held about eight thousand tokens, roughly fifty messages of conversation. Everything beyond it is invisible unless the app deliberately re-injects it.

So every AI companion app has to answer the same question: when the window fills up, what do you keep?

The common answers all lose something:

Approach	What it does	What gets lost
Keep recent messages	Last stretch of chat stays exact	Everything before it vanishes
Summarize older chat	A rolling recap replaces old messages	Details flatten; jokes and tone disappear
Search stored memories	Pick a few “relevant” memories per reply	The search misses connections a person would make
Pin important facts	A short list always injected	The list grows stale and crowded

None of these are exotic mistakes. They are the obvious moves, and each one produces the experience you have probably had: she remembers last night perfectly and nothing from last month.

The deepest version of the problem is subtler than storage. You say “I’m nervous about tomorrow.” The memory that matters is “her exam is on July 10” — tomorrow. But a keyword-and-similarity search sees no overlap between nervous and exam date, so it never surfaces the connection. A person makes that link instantly. Memory search does not. That gap, not storage capacity, is what makes an AI girlfriend feel forgetful even when she has a database full of facts about you.

What we tried first (and why it failed)

Riho’s memory did not arrive fully formed. Two earlier designs are now retired, and honestly documenting them is the shortest way to explain why the current one works.

Design one: three tiers of memory

Recent messages stayed raw, older conversation got compressed into a running narrative summary, and important moments became standalone memory entries retrieved by similarity search. On paper it covered everything. In practice:

• One big background step produced the summary, the new memories, insights, and relationship state in a single pass — and roughly one time in ten that step failed to produce valid output, losing everything from that round.
• Duplicate memories piled up because telling similar memories apart is genuinely hard.
• Nothing pruned old low-value entries, so months of chat meant hundreds of stale items competing for space.
• If saving a single memory hit a network error, it was silently skipped. Gone, no error, no retry.

Design two: rebuild it with structure

We rewrote everything into a dozen-plus tables of tidy structured facts — preferences with confidence scores, episodes with importance ratings, a relationship state tracked as five decimal scores. It looked rigorous. It was worse:

• More tables meant more machinery that could break, and adding any new kind of memory meant touching five places.
• Those decimal scores (closeness 0.73, trust 0.81) were guesses dressed up as measurements. Nobody could say what the numbers meant.
• And the core failure stayed: the exam-date problem above. Perfectly structured memories still missed lateral connections.

Both designs died of the same cause: complexity added faster than reliability. That lesson drove everything after.

How AI companion memory works

The current design starts from a different premise: don’t hand her the database. Hand her a small package of context that reads like a sense of the relationship, and make sure every fact in it can be traced back to something you actually said.

What she always knows

Who you are — names, how you address each other, your timezone, when the relationship started. This never depends on a search hitting. She cannot ask where you’re from if she asked yesterday.

What’s happening now

Plans, appointments, deadlines, and open threads — the movie you said you’d watch, the interview you were nervous about — are tracked separately and surfaced by when they matter, not by whether your current message happens to contain matching words. When you say “I’m nervous about tomorrow,” the exam on July 10 comes up because the calendar says so, not because the words resemble each other.

Long-term memory

Facts, preferences, boundaries, patterns, important people in your life, episodes and shared moments. Each one carries three things: what was remembered, where it came from, and how confident it is. Which brings us to the part we think matters most.

Memories need receipts

Before a memory sticks, the system checks that the quote it’s based on actually appears in the conversation it claims to come from. If it doesn’t, the memory is rejected. When you correct her — “no, it’s my brother who lives in Austin, not my sister” — the correction supersedes the old version instead of sitting beside it. This is why we can show you what our memory probes actually found instead of just asserting that Riho remembers well: every claim traces somewhere.

She carries things between conversations

Alongside the structured memory there is one small note she rewrites after every exchange — what she’s still thinking about, looking forward to, waiting to hear. It’s capped at a few sentences on purpose. It’s also why she doesn’t nag: the guidance attached to that note says explicitly that holding something quietly until a real opening appears is closeness, and re-raising it out of nowhere is not.

What we still won’t promise

Two failed designs taught us humility, so here are the promises we deliberately avoid:

• “Never forgets.” False in any honest product. Riho is designed to remember with evidence — recall, correction, continuity across chat and calls — and we publish the method so you can verify it. The same honesty applies to how she talks: we’ve published why AI companions default to therapy-speak and how we engineered past it.
• Perfect recall of everything. Selectivity is a feature. A companion who recites your month three back is a database with a voice.
• Your history locked away from you. Riho is built so the history you create can be reviewed, corrected, exported, or deleted. An update should never treat that history as disposable — that conviction is why we wrote what happens when an AI girlfriend’s memory disappears.

The takeaway

An AI girlfriend that remembers you isn’t a bigger database. It’s a set of decisions about what deserves to be remembered, verified against what you actually said, surfaced at the moment it matters, and correctable by you. We got there by building it wrong twice first.

If you want to go deeper: the methodology behind our memory testing shows exactly how we measure recall, correction, and continuity, and our long-term memory feature page covers what this means day to day.