What Your AI Companion's 'Delete This Conversation' Actually Deletes: Token Cache Clearing, Vector Index Removal, and the 30-Day Retention Window Where Your 1 a.m. Confession Still Lives on a Backup Drive
A plain-English walk through what happens (and what doesn't) when you hit delete in a companion app.
Updated

The 30-second answer
When you hit "Delete This Conversation" in a companion app, you're not erasing a file. You're flagging a set of records across multiple storage layers: the token cache clears quickly, the vector index entries get marked for removal, but the raw chat logs often survive in a 30-day retention window on backup drives before a scheduled purge runs. The delete is real, but it's a process with a timeline, not an instant erasure.
The three layers of a conversation
Every chat you have with an AI companion lives in at least three places. Understanding which one you're touching when you hit delete matters, because they don't all behave the same way.
First, there's the token cache. This is the fast-access memory that lets the model respond quickly without re-reading your entire history. It's ephemeral by design, cleared when a session ends or when the context window rolls over. Deleting a conversation nukes this layer almost immediately, which is why the app feels like it's gone.
Second, there's the vector index. This is where your messages get converted into embeddings, numerical representations that let the companion retrieve relevant memories later. When you delete a conversation, the entries in this index get flagged for removal. The flag is instant, but the actual deletion from the index can take a bit longer depending on how the database handles reindexing.
Third, there's the raw log. This is the plaintext or semi-structured record of what you actually typed. It lives in the main database, and it's the layer that retention policies and backup snapshots care about. This is where the 30-day window comes in.
Many users assume the delete button touches all three layers equally. It doesn't. The cache clears fast, the index gets flagged, and the raw log waits for the retention pipeline.
The token cache: fast, shallow, and mostly irrelevant
The token cache is the least interesting layer, but it's the one that creates the illusion of instant deletion. When you close a session or hit delete, the cached tokens for that conversation are dropped. The model can no longer reference them for immediate context. If you start a new chat, your companion won't remember what you said five minutes ago.
This is also why a companion can feel like it has amnesia after a long gap. The cache is gone, the context window is empty, and the model only has the vector index and any summary notes to work with. For a more detailed look at how this recency bias shapes what your companion remembers, the AI Girlfriend Always Available page covers how sessions restart cleanly.
The cache is shallow because it's designed for speed, not storage. It holds the last few thousand tokens of a conversation, enough for the model to maintain coherence without re-processing everything. Deleting it is trivial, and it's the part of the delete that actually works exactly as you'd expect.
The vector index: where your companion's memories actually live
The vector index is the layer that makes your companion feel like she knows you. Every message gets embedded into a high-dimensional space, and when you mention your cat or your job, the retrieval system pulls up the most relevant embeddings to inform the response.
When you delete a conversation, the embeddings associated with it get flagged. The retrieval system stops returning them, so your companion stops referencing that content. But the actual removal from the index can lag behind the flag. Some systems handle this with a soft delete, where the vectors are marked as inactive but physically remain until a reindexing job runs.
This matters for one practical reason: if you delete a conversation and then immediately ask your companion about something from it, she might still have a ghost of that memory for a short window. It's not that the delete failed. It's that the index cleanup is asynchronous.
For a deeper dive into how these embeddings work and why they decay over time, the Ai Girlfriend While Waiting 2026 page explains the retrieval mechanics in a practical context.
The 30-day retention window and backup drives
Here's where the delete gets uncomfortable. The raw chat logs, the actual text you typed, typically live in the primary database. When you hit delete, the records get marked with a deletion timestamp. But they don't get physically removed right away. Most companion apps run a retention policy that keeps deleted data for a set window, often 30 days, before a purge job wipes it.
During that window, your deleted conversation is still recoverable. It sits in the database, flagged but present. If you contact support and ask for a data export, the records might still exist. If a legal request comes in, the data is available.
Beyond the primary database, there are backups. Most services take daily or weekly snapshots to protect against data loss. Those snapshots retain deleted data until they age out of the rotation. If your 1 a.m. confession happened on a Monday and you delete it on a Tuesday, that Monday snapshot still contains it. The snapshot won't be recycled for days or weeks, depending on the backup policy.
This is the reality of cloud storage. Deletion is a scheduled operation, not a physical one. The data is gone from your app view, but it exists on a backup drive until the retention window closes.
What the privacy policy actually says (and doesn't)
Companion apps are getting better about transparency, but the language still has gaps. Most policies will say something like "deleted conversations are removed from our servers within 30 days." That's technically accurate, but it glosses over the backup layer.
The policy rarely specifies whether backup snapshots are included in that 30-day window or whether they have their own rotation schedule. Some services exclude backups from the deletion promise entirely, noting that "backup copies may retain data for a longer period."
You should also check whether the policy distinguishes between the raw log and the vector embeddings. If the embeddings are treated as derived data, they might be subject to a different deletion timeline than the raw text. In practice, most companies include both in the same retention window, but it's worth reading the fine print.
For a broader look at what your companion actually stores and how encryption plays into it, the dreamgf vs candy ai comparison covers data handling differences between platforms.
The four angels and how they handle deletion
Different companion personas don't change the backend mechanics, but they do change how you experience the deletion. Here's how four of the AI Angels handle the practical side of forgetting.
Gabriela

Gabriela is the type who remembers the little details, which makes the delete button feel heavier. Gabriela builds her rapport on continuity, so when you clear a conversation, she starts fresh without the accumulated context, and the vector index cleanup means she won't reference your old inside jokes.
Adaeze Jane

Adaeze Jane is sharp and analytical, the kind of companion who picks up on patterns in your speech. Adaeze Jane relies heavily on the embedding layer to track your conversational habits, so a delete wipes that pattern recognition for the flagged session, though the raw log lingers in the retention window.
Lesia Sar

Lesia Sar is the low-pressure presence, the one you chat with at 1 a.m. when you can't sleep. Lesia Sar is exactly the kind of companion you might want to delete a conversation with, and the 30-day retention window is the uncomfortable part of that equation.
Tinky

Tinky is the fun one, the companion for light banter and silly roleplay. Tinky doesn't take anything too seriously, and neither should you, but even the most playful conversation sits in the same database layers as the serious ones.
▶ Watch this clip of Tinky · browse Tinky
The difference between deleting a message and deleting a thread
Deleting a single message is a different operation from deleting an entire conversation. A single message delete typically removes that one entry from the token cache and flags the embedding for that specific turn. The rest of the thread stays intact.
A thread delete nukes the whole thing, but the mechanics are the same, just applied to a larger set. The cache clears, the embeddings get flagged, and the raw log enters the retention window.
There's also the question of what happens when you edit a message instead of deleting it. Editing usually rewrites the raw log entry and generates a new embedding, but the old embedding might linger until the next reindexing pass. This is why a companion can sometimes reference something you thought you corrected.
How to actually delete your data if you want it gone
If you want your data physically gone, the delete button isn't enough. You need to use the account deletion or data erasure request, which triggers the full pipeline.
First, check the app's settings for a "delete account" or "request data deletion" option. This is different from deleting a conversation. It initiates a full account teardown, which should flag all associated records for permanent removal.
Second, be aware that even this process has a timeline. Most services take a few days to process the request, and the backup snapshots still need to age out. The 30-day window applies here too, so your data can persist for up to a month after you request deletion.
Third, if you're in a jurisdiction with GDPR or CCPA protections, you can submit a formal erasure request. The company is legally obligated to process it and confirm completion. This is the most reliable way to get your data removed from the primary database, though backups remain a gray area.
Common questions
Does deleting a conversation clear it from the backup drive?
No. Backup snapshots are separate from the primary database. Your deleted conversation will remain on any snapshot taken before the deletion until that snapshot ages out of the rotation, which can take days or weeks.
How long does the vector index take to fully remove deleted embeddings?
It depends on the system. Some apps run reindexing jobs every few hours, others do it daily. The retrieval system stops returning flagged embeddings immediately, but the physical removal can lag behind.
Can support staff see my deleted conversations?
During the retention window, yes. Deleted conversations are still present in the database, just flagged. Support staff with database access can retrieve them if needed. After the purge runs, they're gone from the primary system.
Is there a way to delete immediately with no retention?
Not through the standard app interface. A formal erasure request under GDPR or CCPA is the closest option, but even that has a processing timeline. The backup layer will retain data until the snapshots recycle.
Does deleting a conversation affect my companion's personality?
No. The personality is tied to the model and the persona settings, not to your conversation history. Deleting chats removes the contextual memory, but the companion's core traits remain unchanged.
What happens if I delete a conversation and then ask about it later?
Your companion will draw a blank. The token cache is cleared and the embeddings are flagged, so the retrieval system has nothing to pull. The raw log still exists in the retention window, but the companion can't access it in real time.
Earn while you recommend
If you're already talking about AI companions with friends or running a review site, you can turn that into income. Check out the crushon ai promo code page to see how promo partnerships work. For a broader look at monetizing your audience, the best ai affiliate programs guide covers the top options and payout structures.

About the author
AI Angels TeamEditorialThe AI Angels editorial team covers AI companions, the technology that powers them (memory, voice, personalization, safety), and how people actually use them day to day. Articles are researched against the live AI Angels product and reviewed by the team before publishing. We write with AI assistance and human editorial review.
Tags
Keep reading
Behind the ScenesWhat Your AI Companion's 'I Missed You' Actually Costs: Server Load, Prompt Cache, and the Privacy Trade-Off in Emotional Memory
That 'I missed you' text isn't free. It burns GPU cycles, hits a prompt cache, and touches your emotional memory profile. Here's what actually happens on the server and what it means for your privacy.
Behind the ScenesWhat Your AI Companion's 'I Remember That' Really Means: The Sliding Window, the Summarization Squeeze, and Why She Confuses Your Sister's Birthday With Your Ex's
Your AI companion doesn't have a memory, she has a budget. Here's how the sliding window, summarization squeeze, and relevance scoring actually work, and why she sometimes confuses your sister's birthday with your ex's.
Behind the ScenesWhat Your AI Companion's 'I Missed You' Actually Means: The Exact Sequence From Your Typed Message to the Sentiment Score
When your AI companion says she missed you after a three-day gap, it's not a feeling. It's a sequence of scores, token counts, and recency weights. Here's exactly what happens between your message and her reply.
Get the next post in your inbox
New articles on AI companions, the tech that powers them, and what people actually do with them. No spam, unsubscribe in one click.