Memory upgrade plan — semantic search with FAISS or ChromaDB

[chansg]

Overview

Documentation issue — planning only, no implementation in this session.

Current state

Aria's memory uses SQLite with episodic and semantic tables.
All conversations are stored but retrieved loosely without relevance ranking.
There is no importance tagging, no decay for old memories, no semantic similarity search.

The problem

After many conversations Aria cannot surface the most relevant past context.
She stores everything but retrieves indiscriminately.
This is the core intelligence bottleneck identified by independent review.

Proposed upgrade

Replace or augment SQLite retrieval with a vector search layer:

• FAISS (local, fast, no extra service) or ChromaDB (persistent, feature-rich)
• Each memory embedded as a vector using a sentence transformer model
• Retrieval ranked by semantic similarity to current query
• Importance tagging: flag high-value memories (user goals, key facts)
• Time decay: recent memories weighted higher than old ones

Why this matters for streaming

During a Valorant stream Aria should recall:

• Chan's preferred agents and playstyle
• Previous match results and commentary style preferences
• Viewer-specific context if it develops over time

Implementation order (future sessions)

1. Add FAISS or ChromaDB alongside existing SQLite
2. Embed new memories on write
3. Retrieve by similarity on read
4. Deprecate flat SQLite retrieval for conversation context

This issue is documentation only — implementation planned for a future session.