Engram tops LoCoMo with no-LLM retrieval

This reads less like a benchmark brag and more like a practical recipe for making chat-memory retrieval stop being dumb. The speaker-name injection result is the most interesting part, because it exposes a common failure mode in first-person conversation logs that standard retrieval stacks miss.

• –Chunking long sessions into smaller overlapping windows preserves fact-level signal instead of smearing it across an entire conversation
• –Prepending timestamps helps both dense and sparse retrieval answer time-scoped questions without relying on brittle metadata filters
• –Injecting speaker names closes the gap between first-person turns and name-referenced questions, which is why multi-hop recall jumps hard
• –The numbers are retrieval-only, so they do not compare cleanly with end-to-end QA F1 claims from other systems
• –The stack is still compute-heavy on CPU, but it is a credible local alternative if you want no API dependency and can tolerate reranking latency