How AI Companion Memory Works
Context windows vs memory journals, what “remembers you” means, and how to test recall before you subscribe.
Memory Marketing vs. Memory Reality
Apps love the word “memory.” In practice, most combine short-term context (recent messages in the prompt) with optional long-term stores (journals, pinned facts, or profile updates). This guide explains the difference.
Context Window vs. Memory Store
A context window only “remembers” what fits in recent chat history—fine for one session, fragile after logout. A memory store persists facts across days. Nomi exposes this via a journal; AISOUL and Kindroid emphasize continuity in testing; Character.AI memory is adequate for casual use but not archival plotting.
The Three-Fact Memory Probe
Our standard test: share three unique facts (e.g., a pet name, workplace, hobby), chat ten more messages, log out, return next day, ask what the bot recalls. Partial credit for two of three; zero of three after advertised long-term memory is a major downgrade. Full methodology in how we review.
Memory Strength by Product Type
Memory leaders: AISOUL, Nomi, Kindroid (after setup), Replika (bond framing). Casual memory: Character.AI, SpicyChat (varies by bot). Workshop: Janitor AI (depends on model and card). See best for memory.
What We Cannot Prove in a Week
Short editorial tests cannot prove month-long retention. We state actual test span in each review. Products change memory backends without notice—re-run your probe after major updates.
FAQ
Do all apps lie about memory? No—but many oversell. Test yourself. Best memory UI? Nomi for transparency; AISOUL for memory + media together. Does Character.AI remember forever? Treat it as session-to-session continuity, not an archival journal.
When Memory Fails (And What To Do)
Contradiction week: Bot invents new facts—reset thread or switch character; may be model update not “gaslighting.” Logout amnesia: Facts vanish after restart—product relies on context window only; downgrade expectations or switch to Nomi/Kindroid. Card/model swaps (Janitor AI): Memory quality follows your API model—document which model passed your probe. Re-run probes after major app updates; vendors change backends without press releases.