AI memory should be useful without becoming a hidden dossier.
Memory improves continuity only when the user can see what the assistant retained, correct it, remove it, and understand whether it remains on the device.
At a glance
Private AI memory is user-controlled context stored for future assistance with a defined purpose, retention rule, and deletion path. It should separate temporary chat context from explicitly saved facts, show the user what was retained, allow correction and removal, and explain whether any memory is sent to a provider. Phos is designed around local memory controls rather than an invisible cloud profile.
How Phos handles it
Three clear routes
Local model
Memory retrieval and local-model prompting can remain on the phone.
Local server
Retrieved memory included in a prompt is sent to the selected personal server.
Bring your own key
Only the memory inserted into a BYOK request reaches the chosen provider, so selection and minimization matter.
Who this is for
Built for people with real private work to do.
People who want continuity without surrendering a long-term behavioral profile.
Users storing preferences, project notes, or recurring instructions on Android.
Privacy reviewers mapping local data, exports, backups, and provider prompts.
Anyone who has seen an assistant confidently remember something wrong.
Plain comparison
The point is control, not a louder chatbot.
These pages are for people comparing real options. Phos should win when someone wants privacy, local control, no account wall, and an assistant that still feels good to use.
| Feature | Phos | Typical cloud chatbot | Raw local app |
|---|---|---|---|
| Saved fact visibility | Designed for inspect, edit, and delete controls. | May be summarized behind account settings. | Often manual system prompts or files. |
| Retention owner | User-controlled local lifecycle. | Vendor policy and account controls. | File owner, if the client exposes state. |
| Provider exposure | Depends on active route and selected recalled context. | Memory normally participates in hosted prompts. | None unless connected to a remote endpoint. |
| Correction | Individual notes can be changed rather than only wiping all data. | Varies by product. | Usually edit files or prompts manually. |
Separate five kinds of memory
Current-turn context, chat history, explicitly saved facts, inferred summaries, and attachments are not the same data. Current context can disappear when the thread closes. Chat history preserves exact words. A saved fact such as a preferred writing style may be intentionally reusable. An inferred summary is riskier because the model may compress nuance incorrectly. Attachments can contain far more sensitive information than either.
An interface should name these categories in normal language. A single 'memory on' switch hides too much: it does not tell the user what will be retained, how retrieval works, or whether the full note enters every prompt. Phos's intended boundary is selective, editable local memory. The safest default is to remember less, retrieve only what helps the current task, and let the user inspect the result.
Retrieval is a disclosure decision
A note can remain safely stored on a phone until the app includes it in a prompt. At that moment, the active inference route matters. A local model can consume the note without a provider request. A local server receives it over the chosen network. A BYOK provider receives whatever recalled context is attached to the request, even if the underlying database remains local.
Good memory systems therefore minimize at retrieval time, not only at storage time. They should prefer the smallest relevant snippet, avoid unrelated personal facts, and make route changes visible. A user asking for a grocery list does not need their startup notes, relationship history, or study profile sent along. Relevance is both a quality feature and a privacy control.
Deletion must address copies and exports
Deleting a visible note should remove the app's active copy and stop it from being retrieved. It may not erase screenshots, exported files, operating-system backups, provider logs from earlier BYOK requests, or copies already sent to a personal server. A trustworthy product explains that boundary rather than promising universal erasure it cannot perform.
Provide three levels: delete one memory, clear a category or conversation, and reset local data. Export should be explicit and readable so the user can review what leaves the app. If backup behavior changes, the product should describe the destination and retention. These controls are more valuable than a decorative privacy badge because they let the user act when circumstances change.
Quality controls are privacy controls
Incorrect memory can repeatedly distort answers. A model may turn 'I am considering moving' into 'I live in a new city,' or infer a durable preference from one temporary request. Letting users edit and reject memory prevents that error from becoming a persistent profile. It also reveals what the system thinks it knows, which reduces the asymmetry between user and assistant.
Useful memory should include provenance: was this saved explicitly, extracted from a particular chat, or generated as a summary? It should support a last-used timestamp and a purpose. Old notes can be reviewed or expired rather than retained forever. The target is not maximum recall; it is accountable continuity, where every retained item has a reason to exist and a clear way to leave.
Direct answers
Frequently asked questions
Can AI memory stay local while using a cloud model?
The database can stay local, but any recalled memory inserted into a cloud request is sent to that provider. Local storage and prompt disclosure are separate boundaries.
Should an assistant automatically remember everything?
No. Selective, purpose-limited memory reduces accidental disclosure, irrelevant context, and persistent mistakes. Explicit saves and inspectable suggestions are safer defaults.
What should a reset remove?
It should remove the app's local chats, saved memory, attachments, and settings covered by the reset. It cannot recall data already exported or sent to another service.