F.R.I.D.A.Y.

Fully Responsive Intelligent Digital Assistant for You

"I've dated holograms with more substance than this, but sure, let's give it a shot."

F.R.I.D.A.Y. is a voice-first AI assistant built on a LiveKit pipeline with a FastMCP backend. She reasons via Google Gemini 2.5 Flash, speaks via the nova TTS voice, and now — she remembers.

---

What's in this repo

This repository contains the Memory System — the first major capability module beyond the base demo.

Module	Description
friday/tools/memory.py	Three MCP tools: remember, recall, forget
friday/agent.py	Voice turn handler with silent auto-recall injection
tests/test_memory.py	14 unit tests covering all memory tools
tests/test_memory_e2e.py	End-to-end 4-turn session simulation

---

How the Memory System Works

On every voice turn, F.R.I.D.A.Y. silently calls recall() with the user's utterance and injects relevant memories into the system prompt — before the LLM ever sees the input. Boss never has to say "remember what I told you."

Boss says something
        ↓
recall(utterance, limit=3)       ← semantic search, silent
        ↓
Relevant memories found?
  YES → prepend to system prompt
  NO  → continue unchanged
        ↓
LLM generates response with full context

The three tools

remember(content, category)   # stores a memory with embedding + dedup check
recall(query, limit)          # semantic search — returns most relevant memories
forget(memory_id)             # deletes a memory by ID

Storage

• Backend: ChromaDB (local, persistent, embedded — no external server)
• Embeddings: sentence-transformers/all-MiniLM-L6-v2 (local, free, fast)
• Deduplication: Cosine similarity threshold at 0.97 — near-identical memories are rejected
• Noise floor: Recall filters results below 0.3 relevance score
• Persistence: Survives process restarts — memories live on disk at friday/memory/chroma_store/

---

Tech Stack

Layer	Tech
Voice pipeline	LiveKit
Tool server	FastMCP
LLM	Google Gemini 2.5 Flash
TTS	nova voice
Vector store	ChromaDB (persistent)
Embeddings	sentence-transformers / all-MiniLM-L6-v2
Tests	pytest + pytest-asyncio

---

Project Structure

F.R.I.D.A.Y/
├── friday/
│   ├── agent.py                  # Voice turn handler + auto-recall hook
│   ├── tools/
│   │   └── memory.py             # remember / recall / forget MCP tools
│   └── memory/
│       └── chroma_store/         # Runtime only — created on first run, gitignored
├── tests/
│   ├── test_memory.py            # Unit tests
│   └── test_memory_e2e.py        # End-to-end session tests
├── requirements.txt
├── CLAUDE.md                     # Claude Code guidance
├── SETUP.md                      # Setup & run instructions
└── README.md

---

Test Results

18 passed in ~105s

tests/test_memory.py         14 passed   ← unit tests
tests/test_memory_e2e.py      4 passed   ← e2e session simulation

---

Roadmap

• get_calendar + set_reminder — act on reminders, not just store them
• Chained tool reasoning — recall → search_web in one autonomous turn
• Voice persona tuning — tighter response formatting for nova TTS
• Memory dashboard — view, edit, and delete what F.R.I.D.A.Y. knows

---

License

MIT