Emotion Machine drops three-tier AI memory architectures

The "filesystem as memory" approach is a clever return to basics that solves the opacity of vector databases while giving agents a literal workspace to manage.

• –pgvector (V1) is discarded as too noisy for nuanced "relationship" memory where semantic similarity doesn't always equal relevance.
• –The "Scratchpad" (V2) offers better control by forcing the LLM to curate its own working context, though it hits token limits quickly.
• –The Filesystem (V3) treats memory as a structured, grep-able directory (e.g., /em/memory/profile.yaml), allowing the agent to use CLI tools for self-management.
• –This transition signals a shift from "retrieval-heavy" RAG to "agent-heavy" memory management where the AI proactively writes and updates its own history.
• –The inclusion of "importance scores" and "hot context" syncing indicates a sophisticated blend of slow and fast memory systems.