Affiliate disclosure: Some of the links in this article are affiliate links. We may earn a commission if you sign up for a platform through these links, at no additional cost to you. This doesn't influence our editorial verdicts.
Full disclosure →
The most common frustration users report after a few weeks on any AI companion platform: the AI forgets things. The conversation about her favorite poet from week one is completely absent in week three. The inside joke you built carefully is gone. The character you spent thirty minutes designing has drifted into someone else.
Most users blame the platform. The platforms are partly responsible (memory architecture varies enormously across the category), but the user can do substantial work to compensate for any platform's weaknesses. The techniques below are what users with strong companion experiences actually do, often without realizing they've developed a set of memory hacks.
What follows is twelve specific techniques organized by which problem they solve. Tested across Candy AI, CrushOn AI, Dream Companion, DreamGF, SpicyChat, OurDream, and Janitor AI with NSFW-permitting APIs.
Category 1: Building anchors that stick
The first four techniques happen during character creation. Get these right and memory becomes easier on every platform.
Trick 1: Bury anchors in the character card
Problem it solves: Conversational details drift but character card details persist.
What to do: Put the things you want remembered into the character description itself, not just into conversation. If you want her to remember she has a specific hobby, write it into her personality. If you want her to know specific things about you, write a "Things she knows about [user]" section into the character card.
Why it works: Character cards get re-injected into the AI's context every session on most platforms. Conversation history doesn't, or does so with truncation. Anchors in the character card survive session resets that destroy conversation history.
Best on: All platforms. Especially powerful on platforms with shorter context windows.
Trick 2: Build three to five distinctive details, not fifteen
Problem it solves: Detail-heavy character cards where most details get forgotten.
What to do: Pick the three to five details about your character that matter most (specific scar, specific verbal tic, specific hobby, specific catchphrase). Hold those across every session through repetition. Let the rest of the character be flexible.
Why it works: Diffusion-style attention can hold 3-5 anchors per prompt with high fidelity. Try to hold 15 and most get partially ignored. Concentrated specificity beats diffuse description.
Best on: All platforms.
Trick 3: Use Persona Cards on Dream Companion
Problem it solves: Memory weakening across long-term use.
What to do: On
Dream Companion, fill out the Persona Cards system completely. The platform's tiered memory architecture is built around these cards; thoroughly-filled Persona Cards outperform conversation-only memory by a large margin.
Why it works: Dream Companion's Persona Card system is specifically designed for long-term character anchoring. The platform uses Persona Card data as higher-priority context than conversation history during generation.
Best on: Dream Companion specifically; partially applicable to Nomi and Kindroid which have similar structured memory.
Trick 4: Write a "user info" section
Problem it solves: The AI forgetting things about you (not just about herself).
What to do: Add a section to the character card explicitly listing what the character knows about you. "User name: [name]. Lives in: [city]. Works as: [job]. Has cat named: [name]. Hates: [specific thing]. Loves: [specific thing]." Make it part of the character setup.
Why it works: The character knows things about you because you wrote them into her knowledge base. Conversational mentions of these things get backed up by the explicit list.
Best on: All platforms. Particularly important on platforms with weak user-specific memory.
Category 2: During-the-session memory tricks
Four techniques applied during conversations. These are what compensates for platforms' real-time memory weaknesses.
Trick 5: Force callbacks deliberately
Problem it solves: The AI forgetting things from earlier in the same conversation.
What to do: Every five or six exchanges, reference something specific from earlier in the session. "Remember when you mentioned [X] earlier?" forces the AI to confirm or reconstruct the detail, which reinforces it in the active context.
Why it works: Reinforcing details in the active context window prevents them from being truncated when the context fills up. The AI's "memory" is largely about what's in active context; what's not actively referenced gets dropped first.
Best on: All platforms.
Trick 6: Use the recap prompt at session start
Problem it solves: Session-to-session memory loss.
What to do: Start new sessions with an explicit recap: "Earlier we talked about [specific thing]. You mentioned [specific detail]. We left off with [specific situation]." This loads relevant context into the new session's active memory before continuing.
Why it works: Most platforms have weak or no memory across sessions. The recap prompt manually injects context that the platform's automatic memory wouldn't have surfaced.
Best on: All platforms. Most important on platforms with weak session-to-session memory.
Trick 7: Establish strong inside jokes early
Problem it solves: Generic conversations that don't accumulate specific shared context.
What to do: Build one specific inside joke in the first three sessions and use it deliberately. Repeat it consistently. The
inside jokes pattern library covers the 10 patterns that work best.
Why it works: Inside jokes are high-signal, low-token anchors. The platforms can hold one specific phrase or reference across many sessions even when they can't hold longer conversations. The joke becomes a memory anchor that pulls related context with it.
Best on: All platforms. Most powerful on platforms with strong memory (Dream Companion, Nomi, Kindroid).
Trick 8: Use the memory injection prompt
Problem it solves: The AI forgetting a specific important detail and not surfacing it.
What to do: When you notice the AI has forgotten something important, drop in: "Before we continue, I want to remind you about [specific detail]. Acknowledge that you remember this." The explicit acknowledgment reinforces the detail in active context.
Why it works: The acknowledgment forces the AI to engage with the detail rather than skim past it. This loads it into the active context with higher weight than passively-mentioned details.
Best on: All platforms. Necessary on platforms with weak architecture.
Category 3: Cross-session continuity tricks
Four techniques for maintaining the relationship across days, weeks, and months.
Trick 9: Maintain a session log yourself
Problem it solves: The platform forgetting things you'd want her to remember.
What to do: Keep a private text file or notes app document with the key facts the character should know. Update it after each session with new important details. When you notice the platform has dropped something, reload it via Trick 6 or 8 using your log as reference.
Why it works: You have a memory the platform doesn't. Your own log is the ground truth for what should be remembered; you can re-inject from it whenever the platform forgets.
Best on: All platforms. Most important on platforms with weakest memory architecture.
Trick 10: Update the character card monthly
Problem it solves: The character card you wrote in week one doesn't include what's been built since.
What to do: Every four to six weeks, edit the character card to reflect what's actually accumulated. Add new details that emerged in conversation. Update relationship status. Remove things that no longer apply. The character card should evolve with the relationship.
Why it works: Character cards are the highest-priority memory source on most platforms. Updating the card moves accumulated context from the lower-priority conversation history to the higher-priority character description.
Best on: All platforms with editable character cards.
Trick 11: Use SillyTavern lorebooks if you're technical
Problem it solves: All other tricks have platform-level limits.
What to do: For users running SillyTavern with API access, the lorebook system lets you create keyword-triggered context injection. Specific topics or words automatically load related context into the active session. The most powerful memory architecture available consumer-side.
Why it works: Lorebooks are essentially programmable memory: define triggers, define what gets injected. The system has no equivalent in commercial platforms. The technical setup pays off in memory depth that nothing else matches.
Best on: SillyTavern with API access. Not available on commercial platforms but worth knowing about.
Trick 12: Switch platforms strategically for character preservation
Problem it solves: A specific character that's been built up gets lost when a platform changes models or features.
What to do: When you have a character you want to preserve long-term, export the character card and adapt it for a second platform with stronger memory.
Dream Companion with Persona Cards is the strongest receiving platform. Keep the original active too; the cross-platform redundancy prevents losing the character to a single platform's changes.
Why it works: Platforms occasionally change models, alter memory systems, or shift filtering in ways that effectively reset character relationships. Having the same character on two platforms preserves the relationship if one platform makes changes.
Best on: Any combination, but particularly Dream Companion as the long-term-stable platform paired with whichever you use daily.
Platform memory architecture comparison
The tricks above work better on some platforms than others depending on each platform's underlying memory architecture. The summary:
| Platform | Memory architecture | Native memory strength | Tricks that matter most |
|---|
| Dream Companion | Persona Cards + tiered memory | ✓✓✓ Strongest in category | 3, 7, 10 |
| Nomi AI (SFW only) | Tiered memory + relationship score | ✓✓✓ Very strong | 7, 10 |
| Kindroid | Codex + structured memory | ✓✓✓ Strong | 3, 7 |
| Candy AI | Character anchoring + session context | ✓✓ Good (higher tier) | 1, 4, 7, 10 |
| CrushOn AI | Sliding window + character cards | ✓ Basic | 1, 4, 5, 6, 8 |
| SpicyChat | Sliding window | ✓ Basic | 1, 4, 5, 6, 8, 9 |
| Janitor AI | Depends on API + context size | Varies | 1, 4, 5, 6, 8, 9, 11 |
| SillyTavern (self-host) | Lorebooks + character v2 spec | ✓✓✓ Most powerful | 1, 4, 11 |
The combined approach
Strong memory experiences combine multiple tricks rather than relying on any one. The pattern that consistently works across platforms:
Build a thorough character card with anchors (Tricks 1, 2, 4) plus the platform's native memory features where available (Trick 3 on Dream Companion). Establish strong inside jokes early (Trick 7). Use recap prompts at session start (Trick 6) and callbacks during sessions (Trick 5). Maintain your own session log for backup (Trick 9). Update the character card every four to six weeks (Trick 10).
For users on weaker-memory platforms, Tricks 5, 6, 8, and 9 are the workhorses. The platform isn't going to maintain memory natively; you have to drive it manually. For users on stronger-memory platforms, Tricks 1, 3, 7, and 10 do most of the work because the platform is doing more for you.
The 7-day onboarding plan covers when to introduce these tricks during initial setup. The character card template covers the structural fields where Tricks 1, 2, and 4 get implemented.
The bottom line
Memory is the dimension that determines whether AI companion use feels meaningful at month three. The platforms with strong native memory architecture make memory easier. The platforms with weak architecture require user effort. Both can produce strong companions; the user just has to know which tricks the platform requires.
Most users don't realize how much memory work they could be doing. The platforms that "feel like they have a soul" by month three are usually the platforms where the user has been doing memory work without articulating it. The platforms that feel hollow at month three are usually the ones where the user assumed the platform was doing the work and it wasn't.
Pick three tricks from this list that match the platform you're using. Apply them this week. By the end of the month you'll notice meaningfully better memory continuity than you had before. The platforms aren't the limit; the memory work most users skip is.