polymnemo

Store a memory for later semantic recall. Large content is split into smaller chunks on write (one vector each), so this may create several ids. `namespace` groups memories (defaults to the shared namespace); `tags`/`source` are optional metadata. Returns `{ids, chunks, namespace}`.

Search memories by meaning and return the closest matches. Results are ranked by similarity and bounded by `limit` (default 8). Use the returned `next_cursor` with `has_more` to page further — you decide whether the results are enough. `namespace` selects the collection (defaults to the shared namespace). Returns `{items, total, has_more, next_cursor}`.

List stored memories newest-first (no search query). `namespace` selects the collection (defaults to the shared namespace). Page with `next_cursor` / `has_more`. Returns `{items, total, has_more, next_cursor}`.

Fetch a single memory by its id (from `remember`/`recall`/`list_memories`).

Replace a memory's content (re-embeds it). Returns the updated memory.

Delete a memory by id. Returns `{id, deleted}` (deleted=false if absent).

Persist a session's full content under `session_id` for later reload. Content is stored as ordered, losslessly-reassemblable chunks (also embedded, so it's searchable via `recall`). Re-saving the same `session_id` replaces it. Returns `{session_id, chunks, chars, namespace}`.

Reload a saved session's content, one character page at a time. Page with `page` (0-based) while `has_more` is true. Returns `{session_id, content, page, page_size, total_chars, has_more}`.

Register a file / image / video memory and get a URL to upload its bytes. The bytes never go through this channel: `description` is embedded so the file is findable via `recall`. **PUT** the raw bytes to `upload_url` sending `upload_headers` (the signed Content-Type). **Then call `confirm_upload`** — the memory stays hidden from `recall` (and the size cap is enforced) until you do. Media defaults to a private namespace. Returns `{memory_id, object_key, upload_url, upload_headers, content_type, namespace}`.

Confirm a media upload after you've PUT the bytes. Verifies the object exists, records its real size + checksum (and rejects an over-limit upload), then makes the memory findable via `recall` / downloadable. `id` is from `create_upload`. Returns `{memory_id, confirmed, size_bytes, content_type}`.

Get a short-lived URL to download a media memory's bytes. `id` is the media memory's id (from `create_upload` / `recall`). Returns `{memory_id, url, content_type}`.