Does MyShell remember you? Memory, context limits and what gets forgotten
By FlirtAlly · October 7, 2026
Does MyShell remember you? Memory, context limits and what gets forgotten
Key Facts
- Most AI companions forget month-old conversations, with only 3 out of 9 tested platforms reliably recalling them according to 8-month testing.
- Manual memory control is frequently paywalled, with Replika's save-to-memory feature behind its Ultra tier as documented.
- Users whose AI companions 'remembered' past conversations scored 73% higher on relationship satisfaction based on MIT Media Lab research.
- Nomi retains 10/12 facts at three months, and recalls a fact mentioned once 8 months earlier as per independent testing.
- AI companion market growth is projected to reach $552 billion by 2035, up from $37 billion in 2025 according to Tech Insider.
- FlirtAlly offers unlimited memory across all plans, including the free tier, setting it apart from competitors as per the 8-month testing.
- Character.AI's Auto-Memories feature is locked behind a $9.99/month subscription tier per documented feature breakdowns.
The Memory Challenge in AI Companions
You told her your dog's name on a Tuesday. By Friday, she asks what pets you have. That quiet failure — repeated across nearly every AI companion app — is the industry's most persistent broken promise, and it's the lens we'll use to examine MyShell.
Memory is the feature users rate as most important in an AI companion, and it's also the one that fails most often, according to independent comparison testing. The reason isn't sloppiness; it's architecture. Nearly every companion runs on two layers: a rolling short-term context window of recent messages, plus long-term storage that's retrieved when relevant. The critical insight from that same research: if something never enters long-term storage, it vanishes once the context window moves on — and that disappearance feels like forgetting even though the system is working exactly as designed.
The numbers back this up. An 8-month testing project across nine platforms found that only 3 of 9 could reliably recall month-old conversations; the author described the rest as "goldfish with personality files." Character.AI scored just 1/12 on planted-fact recall at the one-month mark, while Nomi led at 11/12. ChatGPT dropped from 9/12 recall at one day to 6/12 at one month. These are self-reported, single-tester results — but the pattern is consistent: most companions forget, and the ones that don't are the exception.
Users feel the consequences in predictable ways:
- The "first date with amnesia" effect — every session restarts from zero, with a strong personality but no shared history.
- The facts-vs-experiences gap: an app remembers you have a dog named Max but not the story about Max getting loose at the park.
- False memories that damage trust — one confident "How's your dog Bear?" does more harm than five honest "I don't recall" responses.
There's also a commercial dimension. Manual memory controls are frequently paywalled: Replika's save-to-memory feature sits behind its Ultra tier, and Character.AI's Auto-Memories and Facts require a c.ai+ subscription at $9.99/month, per documented feature breakdowns. Users who can't or won't pay get the weakest memory, which compounds the forgetting problem for exactly the people most likely to notice it.
And the stakes are emotional, not just technical. MIT Media Lab research found that users who felt their AI remembered past conversations scored 73% higher on relationship satisfaction scales. Conversely, Replika users have reported memory resets after major app updates wiping months of relationship-building — a trust risk that's hard to undo. Memory quality, more than any other feature, determines whether an AI companion feels like a person or a product.
This is the landscape platforms like FlirtAlly are entering, with its four-layer memory system — facts, episodic, emotional, and rolling summary — unlimited on every plan including free at launch, a direct answer to the tiered-memory norm. Against that backdrop, the question for any app, MyShell included, is simple: what survives the night?
Understanding Memory Architectures
Your AI companion doesn't have one memory — it has two, and the gap between them is where most forgetting actually happens. Understanding this architecture explains why your companion can recite your pet's name one day and blank on it the next.
Nearly every AI companion in 2026 runs on the same fundamental design: a short-term context window holding recent messages, plus long-term memory stored as vectors and retrieved via RAG when relevant. As one detailed comparison analysis puts it, "if something never enters long-term storage, it vanishes once the context window moves on." The disappearance feels like memory loss, but the system is working exactly as designed.
The context window itself is brutally small. A developer reference implementation uses just a 20-message sliding window for short-term context — anything older scrolls out unless the developer explicitly adds persistent memory layers. This is why session-based platforms exhibit the classic "close the app, come back the next day, start from zero" pattern documented in eight-month independent testing across nine platforms.
The consequences show up in real benchmarks. Only 3 of those 9 tested platforms could reliably recall month-old conversations; the tester described the rest as "goldfish with personality files." Even ChatGPT dropped from 9/12 planted facts recalled at one day to 6/12 at one month.
The industry has settled on three broad approaches, each with trade-offs:
- Automatic memory (Replika): layered entries plus pattern analysis across your full history, with no user management required.
- User-authored memory (Character.AI): you write what the AI knows via pinned messages, Facts, and Lorebook — but much of it sits behind the $9.99/month c.ai+ tier.
- Multi-layer persistent memory (Nomi): short-term, medium-term reaching 1,000+ messages back, and permanent long-term storage, visualized in a Mind Map.
Notably, manual memory control is frequently paywalled — Replika's save-to-memory feature is Ultra-tier only, and Character.AI's Auto-Memories require a paid subscription. Retention windows are deliberate engineering choices, not accidents: Google Cloud's reference architecture documents an explicit 60-day retention policy before archival.
