Lossless Claw preserves full agent memory

This is the right kind of memory layer for long-running agents: keep everything, compress aggressively, and make retrieval first-class instead of pretending context windows are storage.

• –Saves every message locally before compaction, which is a big step up from lossy summarizers that silently drop details.
• –DAG summaries plus `lcm_grep`, `lcm_describe`, and `lcm_expand` give agents a real retrieval path back to source messages.
• –The local SQLite design is privacy-friendly, but it also means users now own backup, encryption, and retention hygiene.
• –Quality still depends on the summarizer model and compaction thresholds, so this is infrastructure, not magic.