Emotion Machine drops three memory architectures for AI

The shift from vector search (V1) to a structured scratchpad (V2) and agent-managed filesystem (V3) reflects a broader industry move away from "dumb" retrieval toward agentic curation.

• –Vector search is excellent for facts but misses the nuance of personality and evolving relationship dynamics.
• –The "Scratchpad" model (V2) mimics human short-term memory by keeping high-priority context permanently in the prompt, but eventually hits context window limits.
• –The "Filesystem" approach (V3) is the most ambitious, treating the AI as an agent with its own workspace, using standard tools like grep and cat to manage massive context.
• –Materializing context as real files enables easier debugging and multi-agent collaboration compared to opaque vector blobs.
• –"Hot Context" caching solves the latency-vs-depth trade-off, enabling millisecond chat responses while maintaining gigabytes of potential history in the filesystem.