ADR 0001 — MemoryOps is a governance layer, not a memory database

Status

Accepted draft.

Context

AI agents need memory across sessions, tools, and projects. Existing approaches often focus on vector storage or raw chat history. This creates memory pollution, stale facts, privacy risk, and poor skill quality.

EverOS demonstrates a strong local-first memory runtime architecture. However, MemoryOps is intended to be reusable across many runtimes and backends.

Decision

MemoryOps will be designed as a governance/control plane for agent memory and learning.

It will own:

• classification;
• policy decisions;
• candidate review;
• routing;
• audit;
• skill quality gates;
• lifecycle and expiration;
• backend adapters.

It will not be tied to one storage backend.

Consequences

• The first implementation can be thin and policy-first.
• EverOS can be integrated as a backend adapter, not treated as the whole product.
• CEO AI OS remains the first customer but not the product boundary.
• GitHub Issues/PRs become the coordination surface for multi-agent development.