Design Philosophy

Text-First

Long-horizon history should remain readable and portable. Mems uses JSONL for L3 archive output instead of locking everything into one vector store or proprietary format.

Decentralized Schema

Mems does not assume one fixed upstream identity model. tenant_id, user_id, agent_id, session_id, and scope together define memory isolation and visibility.

Memory, Not Runtime

Mems stores memories, not agent runtime state. It does not manage session context, task goals, active plans, or temporary variables. The caller owns its own runtime.

Layered Memory

Hot cache, raw events, stable knowledge, and cold archive are not the same kind of data. Mems keeps them in explicit layers instead of collapsing them into one giant memory bucket.

SQL-First Online Truth

Current online memory is committed to SQL first. Qdrant is treated as a derived online replica, while JSONL is reserved for L3 archive output.

Model-Evolving Memory

Long-term memory should improve as models improve. L3 JSONL is kept as durable, model-agnostic feedstock; a manual re-distillation (POST /v1/mems/redistill) sends archived records back through the current distillation pipeline, and the L2 version chain upgrades old knowledge non-destructively. Mems is designed so the archive is not a tombstone but a source that can be reborn under a newer, stronger model.