Does context survive across sessions?

Long-running agent work does not fit inside a single conversation. Context windows fill, sessions end, machines restart, and the next run has to pick up where the last one stopped. CueCrux treats continuity as a first-class mechanism rather than something you approximate by pasting old transcripts.

The unit of work is the session. A session accumulates facts, decisions and progress as structured, receipted state in the daemon rather than as chat scrollback. Checkpoints let you pin a known-good state at any point. Cross-session handoffs bundle the relevant facts and work items so a later session, or a different agent entirely, resumes with the same grounding. For moving context between machines or teams, .cruxpack export and import carries a portable, verifiable snapshot.

ExecPlans and the live work board hold the longer arc: milestones, gate results and decision records live in the store, so resuming work means reading state, not re-deriving it from prose.

The evidence for doing it this way is public. In ScoreCrux, our open benchmark, stuffed-context recall collapses at 2M-token scale, falling from 44% to 28% on one model and from 28% to 8% on another, while tool-mediated memory holds 80-100% on the same tasks. Answer quality tells the same story: near-identical scores at a fraction of the context tokens.

Continuity through structured, budgeted retrieval is cheaper and more reliable than continuity through ever-longer context. That is the design bet, and the numbers back it.