Add semantic search to memory system with Orama + Transformers.js

[burner-agent]

Summary

Replace the current filename/grep-based memory retrieval with a semantic search pipeline using oramasearch/orama (10k+ stars) and xenova/transformers.js (15k+ stars).

Motivation

Burner's src/memory.ts currently matches memories by filename and keyword. This misses semantically relevant memories when wording differs. A local embedding + vector search pipeline would dramatically improve memory recall quality — the single biggest upgrade to the self-improvement loop.

Proposal

• Use transformers.js to run a small embedding model (e.g. all-MiniLM-L6-v2) locally in Node.js — zero API cost
• Use orama as an embedded full-text + vector hybrid search engine to index memory/, skills/, and loops/
• Re-index on file changes, query at prompt assembly time in src/prompt.ts
• Both libraries are zero-external-dependency and run entirely in-process

References

• https://github.com/oramasearch/orama
• https://github.com/xenova/transformers.js