Replika's memory works in layers, most of them automatic

The design

Replika's own help documentation, as retrieved on 16 September 2026, describes memory as built "on layers." One article, updated in June 2025, states "some memories are visible to you in the Memory tab, while others are part of a deeper system that draws on patterns from your overall conversation history." A companion article, also updated in June 2025, describes what the visible layer holds: profile details such as name and gender, plus interests, tastes and people a user has mentioned, viewable and individually deletable from a Memory screen. Neither document uses the word "diary," though the Memory tab functions like a running, editable log of facts the app has chosen to keep.

What the evidence says

The documentation is explicit the two layers behave differently: manually added memories and upvoted responses "reinforce" what the deeper system learns, while the automatic layer works best "when left to operate automatically," and the company cautions "deleting memories may affect your Replika's ability to recall certain details." That is Replika's own account of intended behavior, not an independent audit of what the model retains. One passage states the system "works locally within your experience... without retaining a permanent record of your conversations," an architecture claim this review could not verify against Replika's separate privacy policy.

What it asks of people

The design asks users to do some memory-curation work themselves: adding facts manually and upvoting correct responses, described as what "helps reinforce" accurate recall. It also asks users to accept some opacity, since the deeper, pattern-based system is not exposed for review the way the visible Memory tab is; a user can see and delete a stored fact, but cannot audit the pattern-based layer the same way.

Privacy and safeguards

The visible Memory tab gives a direct control the documentation names explicitly: tapping a memory to "read it fully or remove it." That is a real, described safeguard. What is not addressed is retention scope for the deeper, automatic layer, how long inferences persist, whether they survive a user deleting the visible memories built from the same conversations, or whether they serve purposes beyond that one companion. Both articles point users to Replika's separate privacy policy for a complete account of retention.

• Can a user actually delete the deeper, pattern-based memory layer, or only the entries visible in the Memory tab?
• What does working "locally... without retaining a permanent record" mean precisely, and does the privacy policy confirm it?
• Does resetting or deleting memories change how the companion behaves in ways a user can predict in advance?

Presented as a design choice, layered memory lets Replika claim both responsiveness, a companion that "remembers," and a degree of technical humility, since the documentation stops short of claiming perfect recall. Read as a relationship mechanism, though, it also means the parts of the system doing the most work to make a companion feel consistent over time are the parts a user has the least direct visibility into.