You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Advice on shared memory for ChatGPT mobile + private Android assistant (self-hosted?) #2106
I'm exploring an opt-in shared memory for two assistants: ChatGPT and a private Android/web AI assistant. We've built a small prototype with fact provenance, revisions, cancellation/supersession safeguards, and an offline-first direction. Our priority is a working end-to-end conversation → save → new session → retrieve flow, not just benchmark numbers.
Your project looks especially promising because of local Markdown storage, MCP, and mobile/cloud sync. Could you please clarify:
What is the currently supported, tested route for querying and updating the same Basic Memory knowledge base from the native ChatGPT Android app (as distinct from ChatGPT web / Custom GPT actions)? Are there mobile-specific limitations?
Could a separate private Android app use your documented HTTP/API interface to share the same memory while keeping user approval and provenance? Which docs/examples should we start with?
For a budget-conscious single user, is a self-hosted remote endpoint realistic, or is Basic Memory Cloud necessary for cross-device access? Could the Android app keep a safe offline cache and sync later?
How are contradictory or superseded facts handled so outdated assertions don't silently resurface as current truths?
Pointers to existing documentation, examples, or related issues are very welcome; no need for custom development. We won't share private code or data publicly. Thanks for building this!
Hi again — a short clarification in case my original question was too broad. I noticed the automated Claude Issue Triage workflow failed; I realize that may be a workflow issue, not a problem with the question.
Our one core goal: have one user-controlled memory store shared by two assistants:
Assistant A: ChatGPT on an Android phone.
Assistant B: our separate private Android/web app.
A person tells either assistant a fact; it is saved with consent and source information; after starting a new conversation the other assistant can recall the current fact. Old/retracted facts must not silently return. Ideally, the Android app could use a local cache offline and sync later.
The specific thing we're asking: Does Basic Memory already support a working setup for this, and if so, which documented client/API/connection path should we try? In particular, please distinguish (1) ChatGPT's native Android app from ChatGPT web, and (2) self-hosted free Basic Memory from the paid Cloud service.
A simple yes / partially / not yet, with a link to an example or relevant documentation, would be hugely helpful. We aren't asking for custom development, and won't share private data or credentials. Thank you!
ChatGPT on Android: yes. ChatGPT reaches Basic Memory through a remote MCP connector, and connectors work in the Android app as well as on the web. The supported route is the Basic Memory Cloud app in the ChatGPT directory.
Your own app: yes. Your app can connect to the same MCP server ChatGPT uses (bm_search, bm_read, bm_write, bm_edit). Use that rather than the internal HTTP API, which isn't a stable public interface. Asking the user before saving and recording where a fact came from would be up to your app; Basic Memory doesn't do either itself.
Self-hosting: possible. See docs/Docker.md and the Dockerfile / docker-compose.yml in this repo; bm mcp --transport streamable-http serves MCP over HTTP. The self-hosted server has no authentication. For ChatGPT to reach it, it has to be public over HTTPS, so you would need your own auth proxy in front of it. Cloud provides that hosting and auth. There is no offline cache that syncs later for a third-party app. Notes are plain Markdown files, so syncing them (for example with git) is up to you.
Superseded facts: nothing marks them automatically. Notes are plain Markdown that you edit in place, so the way to retire a fact is to edit or remove it, or mark it with a status field in the note's frontmatter and filter on that when searching.
Closing as answered. Feel free to reopen with specifics once you've tried a setup.
A correction to point 4. An observation can carry a temporal qualifier that says when the fact holds:
-[decision]@effective[2026-06-10,2026-07-27) The cache layer will use Redis.
-[decision]@effective:2026-07-27 The cache layer will use Memcached.
A range is half-open: it starts on the first date and ends before the second. A single date means the fact holds from that day on. The kinds are effective, valid, occurred, due and mentioned.
bm_search can filter on these. valid_at="2026-07-28" returns only facts whose range contains that date, so the Redis line above is excluded and the Memcached line is returned. valid_overlaps and time_kind filter the same way. Facts with no qualifier make no claim about when they hold, so a valid-time filter leaves them out.
The full syntax is in bm_man(page="bm-search(3)"), under "Authored valid time".
Activity
gitanas316-stack commentedon Nov 12, 2026
Hi again — a short clarification in case my original question was too broad. I noticed the automated Claude Issue Triage workflow failed; I realize that may be a workflow issue, not a problem with the question.
Our one core goal: have one user-controlled memory store shared by two assistants:
The specific thing we're asking: Does Basic Memory already support a working setup for this, and if so, which documented client/API/connection path should we try? In particular, please distinguish (1) ChatGPT's native Android app from ChatGPT web, and (2) self-hosted free Basic Memory from the paid Cloud service.
A simple yes / partially / not yet, with a link to an example or relevant documentation, would be hugely helpful. We aren't asking for custom development, and won't share private data or credentials. Thank you!
phernandez commentedon Nov 12, 2026
Short answers:
bm_search,bm_read,bm_write,bm_edit). Use that rather than the internal HTTP API, which isn't a stable public interface. Asking the user before saving and recording where a fact came from would be up to your app; Basic Memory doesn't do either itself.docs/Docker.mdand theDockerfile/docker-compose.ymlin this repo;bm mcp --transport streamable-httpserves MCP over HTTP. The self-hosted server has no authentication. For ChatGPT to reach it, it has to be public over HTTPS, so you would need your own auth proxy in front of it. Cloud provides that hosting and auth. There is no offline cache that syncs later for a third-party app. Notes are plain Markdown files, so syncing them (for example with git) is up to you.Closing as answered. Feel free to reopen with specifics once you've tried a setup.
phernandez commentedon Nov 12, 2026
A correction to point 4. An observation can carry a temporal qualifier that says when the fact holds:
A range is half-open: it starts on the first date and ends before the second. A single date means the fact holds from that day on. The kinds are
effective,valid,occurred,dueandmentioned.bm_searchcan filter on these.valid_at="2026-07-28"returns only facts whose range contains that date, so the Redis line above is excluded and the Memcached line is returned.valid_overlapsandtime_kindfilter the same way. Facts with no qualifier make no claim about when they hold, so a valid-time filter leaves them out.The full syntax is in
bm_man(page="bm-search(3)"), under "Authored valid time".