8cbd55c9632aaf95ea7b8741c7ab4909c58492b9
Caching/batching/prefixing moved into Embedder so FakeEmbedder exercises the same cache the Bifrost client uses. Rebuilding an index with an unchanged model embeds nothing; a dim or model change drops the vectors and re-embeds. A token-free query short-circuits before the embedding call.
mdmem
Markdown-canonical agent memory service. See PLAN.md for the full design.
Languages
Python
92.4%
HTML
3.7%
CSS
3.6%
Dockerfile
0.3%