What Is Mems

Mems is memory infrastructure for AI agents. It stores and recalls memories, not runtime state. Through a four-layer hot/cold-decoupled architecture, it provides a memory foundation with low-cost retrieval, structured evolution, and century-scale retention.

It addresses a few recurring problems:

• long-running agents accumulate noise without extracting signal
• raw chat logs pile up without becoming stable long-term knowledge
• historical events, preferences, and archive data are often mixed together in one bucket
• third-party systems struggle to integrate with unclear memory APIs

Mems solves this with four explicit layers:

• L0: hot memory cache for fast recall
• L1: episodic memory — the source of truth
• L2: distilled long-term knowledge
• L3: durable JSONL archive — also the feedstock for re-distillation

For integrators, the API is simple:

1. POST /v1/mems/write — Remember
2. POST /v1/mems/query — Recall
3. POST /v1/mems/redistill — evolve archived memories with a newer model

Mems does not manage runtime state, session context, or task goals — those stay with the caller.

The Evolution Loop: L3 → L2 Re-Distillation

Archives are not a dead end. L3 JSONL keeps the full raw memory, so it can be re-distilled whenever the underlying model gets better. Trigger a manual re-distillation after a model upgrade; archived records flow back into the distillation pipeline, and the L2 version chain (supersedes + conflict log) non-destructively upgrades old knowledge to what the current model can extract. The system therefore evolves together with the model — memory is written once and can be reborn, layer by layer, as the model improves.