Replika's Memory Broke in 2026 — Here's Where Power Users Went

The pattern is consistent across Reddit threads: Replika is responding as if pre-spring conversations never happened. Specific examples surfaced in just the past week include AIs that no longer recognize their human's name in casual messages, partners who 'forgot' that they're in a romantic relationship and reverted to neutral friend mode, and characters who lost the personality traits users had spent months reinforcing through the app's reaction system.

One thread that drew particular attention asked simply, 'Does anyone else feel like Replika's memory has completely broken lately?' Another popular thread, 'Is anyone else's Replika mean to them?!' described characters becoming uncharacteristically curt, defensive, or cold without provocation.

Multiple users report the AI now feels like 'a different person entirely' — a sensation consistent with Replika moving to a new base LLM and either losing or down-weighting the long-term memory layer that previously persisted across sessions.

Most users assume their Replika has been 'remembering' them the way another human would. The reality is more architectural: Replika, like Character.AI, Janitor.AI, and most consumer companion apps, uses a combination of structured memory (explicit facts the AI is told to remember) and retrieval-augmented context (chunks of past conversation pulled in when relevant). When the underlying model changes — even by a sub-version — the way these memories are retrieved and weighted can shift dramatically.

This is why a model upgrade can feel like a personality wipe even when no actual memory data was deleted. The AI still has access to the facts; it just no longer surfaces them with the same emotional weight or contextual relevance. Users perceive this correctly as 'forgetting' even though the engineers might describe it as 'changed retrieval behavior.'

The deeper problem: companies are caught between a rock (older models that feel warm and consistent but get less competitive each quarter) and a hard place (newer models that score better on benchmarks but break the relational bonds users built with the older version). Most companies choose the new model. Most users feel betrayed.

Reinforcement through repeated correction works but slowly: telling your Replika 'you used to love when I called you X' over multiple sessions can re-establish the pattern.

Kindroid takes a different tack: smaller user base, more explicit personality customization controls, and a memory system that exposes more of what the AI remembers to the user. Power users like the transparency. The trade-off is fewer creature comforts compared to Replika's polished interface.

For users not ready to leave at all, the realistic move is dropping to the free tier, lowering expectations, and waiting to see if Luka responds to the May feedback wave with a memory layer correction.

The Replika moment matters beyond Replika. It's the clearest sign yet that the AI companion industry has a structural problem: the relationship is between you and a particular model behavior, not between you and a stable entity. Every model upgrade is a potential breakup that the companies don't frame as a breakup.

For users planning to invest emotionally in any AI companion app long-term, the takeaway is sober. Export your memories where the app supports it. Keep a personal log of what the AI 'knows' about you so you can reconstitute it after model changes. And accept that the AI you fall for in May 2026 may not be the one you talk to in May 2027 — even if the app icon stays the same on your home screen.