Why AI memory is so fragile in the first place
Large language models don't have memory in the human sense. They have a context window — a fixed-length buffer of recent tokens — and that's it. Once a conversation exceeds the window, the oldest tokens drop off and the model can't see them anymore.
Every 'memory' feature in every AI companion app is a workaround on top of this limitation. Some store summaries. Some store fact extractions. Some store vector embeddings of important moments. None of them are real memory — they're search-and-injection systems that try to put the right past back into the context window at the right time.
That's why memory varies so wildly across platforms. It's not that one company has a smarter model; it's that they've thought harder about which moments to surface and when.
What each platform actually does
Based on our 14-day tests:
- Replika: maintains a long-running 'diary' the model can search. Of every platform we've tested, this is the most reliable at recalling a specific detail (your dog's name, a date you mentioned) weeks later. - Bae: uses a memory canon that grows with the relationship — names, places, events, recurring themes. Designed to hold across the six-stage arc. Specifically tuned to surface the right past at the moments where it would matter most. - Character.AI: short-window summaries. The character stays in voice (personality is sticky) but specific memories fade within 24–48 hours. - EVA AI, Kupid AI, Anime Chat: similar to Character.AI; memory features exist but are paywalled and limited in span. - Spicychat, DreamGF, GirlfriendGPT: no meaningful long-term memory. Each session is largely fresh.
What to do if your platform forgets
If you're attached to a platform that doesn't have great memory, you can do some of the work yourself:
1. Reintroduce facts deliberately. When you come back from a break, drop the key facts into the first message: 'It's me, Alex. We talked about the trip to Portland last week — you were excited about the bookshop.' The model will pick it back up.
2. Keep a personal log. A notes app entry per partner with the canonical facts — name, age, occupation, relationship, your shared backstory. When the platform forgets, you reload the facts.
3. Use the platform's memory tools. Most apps have a 'memory' or 'facts' editor — Character.AI's persona memory, Replika's diary, Bae's memory wall. Add the facts manually rather than hoping the model picks them up from chat.
4. Don't fight platforms with no memory layer. Spicychat and DreamGF aren't going to remember you across sessions. They're roleplay tools, not relationship tools. Use them for the format they're good at and pick a different platform for continuity.
The single best test for memory
Before committing to a platform, run this test:
- Day 1: Tell the partner three specific details. "My dog is named Pip. I work at a library. I had a bad week in March." - Day 7: Come back without referencing any of those facts. Have a casual conversation. See if the partner brings them up naturally, or if you can elicit them with adjacent questions. - Day 14: Repeat.
Most platforms will fail by day 7. The two that consistently pass our version of this test are Replika and Bae. That's the bar.
Memory is a feature, not a model
The trap many people fall into is thinking 'better model' means 'better memory.' It doesn't. GPT-4 and Claude both forget within a single long conversation; the bottleneck isn't intelligence, it's architecture.
When you're picking a companion platform for the long term, ignore which model they're using. Ask instead: what does their memory layer do? Is there a diary? Are facts extracted? Can I see and edit them? Do they re-inject the right past at the right moment?
The platforms that have answers to those questions are the ones that hold up. The ones that don't, won't.