Planned

Features we intend to build.

Feature	Status	Notes
Ground image extractions in the graph	Planned	Today an image yields only rows in gb_image_extractions / gb_image_text_blocks: no entities, edges or vectors, and query does not see it. The extraction tables are the input. See images.md.
Ontologies used by remember and query	Planned	Ontologies are loaded and stored as a graph of their own, but neither endpoint uses them yet. See ontologies.md.
query uses timeseries	Planned	query does not read stored series, and treats series_observation artifacts like any other. See timeseries.md.
Fetch timeseries from the standalone endpoint	Planned	Fetching from providers is in remember only; a series is not refreshed automatically. See timeseries.md.
Lookahead protection for timeseries	Planned	A series that runs past a text's reference_time is stored whole, so a memory of an event can contain its future. Clip fetched windows, or label the post-event part. See timeseries-design.md.
Web search for image requests	Planned	web_search is rejected (400) together with image, since the image pipeline extracts no entities yet. Depends on grounding image extractions. See web-search.md.

Considering

Undecided: left out for now, with the reasoning written down.

Feature	Status	Notes
Session memory (per-session turn cache)	Considering	Left out for now; the reasoning, and the least invasive design if revisited, are in not-included.md.