Research
Concepts and open questions behind Venkai, written down as they're worked out — not a claim that any of this is settled or fully published.
What it means to treat memory as infrastructure rather than a feature: a persistent, relational representation of an organization's knowledge, served to any AI system rather than embedded inside one. The distinction this page and /architecture/ both draw on: infrastructure exists once and is queried; a feature is rebuilt inside every product that wants it.
Where similarity-based retrieval structurally cannot answer a question — because the answer is a relationship between documents, not a passage inside one — and what a relational, graph-based representation adds instead. Detailed in the journal article of the same theme, with the specific failure modes retrieval exhibits.
Reframing 'context' away from prompt engineering and toward systems design: a model's context window is assembled input, not accumulated knowledge, and the actual engineering problem is deciding — correctly, and current as of right now — which tokens belong in that window for a given request.
The one item on this page with a number behind it. Token cost of a structural edit versus a full-symbol rewrite, and an in-progress run inside the official SWE-bench harness — reported on /audit/ and in the build log, with the command that reproduces each figure.
Only /audit/ and the build log (/journal/) carry measured numbers. Everything on this page is a concept in progress unless it links to one of those two.