Session intelligence is temporary.
A model can perform excellent work and still lose most of its operational value when the session ends. Decisions, constraints, failed approaches, architecture, and project state disappear unless another layer preserves them. Reconstructing that context repeatedly consumes both human attention and machine time.
Persistence is not the same as accumulation.
Saving everything creates a different failure mode: retrieval noise. Durable memory needs selection. Some information belongs in temporary working context. Some belongs to a project. Some becomes an architectural decision or long-lived operating rule. Other material should expire. The goal is not maximum retention; it is maximum future usefulness.
Memory should be linked to decisions and systems.
Codex Mekanika is designed around the idea that knowledge becomes more useful when it can be connected to the systems, decisions, research, and execution it informs. A constraint should be retrievable when a future route encounters the same boundary. A prior failure should change how the next attempt is planned.
Forgetting is part of the design.
A memory layer that never demotes stale information eventually makes old context compete with current truth. Useful persistence therefore requires review, confidence, provenance, and archival behavior. A system should remember enough to compound learning while remaining capable of updating what it believes.