feat: connect Shane voice sessions to memory

agent/agent.py

@@ -9,6 +9,7 @@ import asyncio
 import json
 import logging
 import os
+import re
 import textwrap
 
 from dotenv import load_dotenv
@@ -41,11 +42,35 @@ GEMMA_API_KEY = os.environ.get("GEMMA_API_KEY", "not-needed")
 WEB_MCP_ENABLED = os.environ.get("WEB_MCP_ENABLED", "true").lower() in ("1", "true", "yes")
 FIRECRAWL_BASE = os.environ.get("FIRECRAWL_BASE", "http://192.168.86.2:3002")
 SKILLS_DIR = os.environ.get("SKILLS_DIR", "/skills")
+VOICE_USERS_FILE = os.environ.get("VOICE_USERS_FILE", "/etc/hope/users.json")
+
+
+def load_user_profile(user_id: str) -> dict:
+    """Load one server-side user profile without exposing its credentials."""
+    with open(VOICE_USERS_FILE, encoding="utf-8") as users_file:
+        registry = json.load(users_file)
+    profile = registry.get("users", {}).get(user_id)
+    if not isinstance(profile, dict):
+        raise ValueError(f"unknown voice user: {user_id}")
+    memory = profile.get("memory")
+    if not isinstance(memory, dict) or not memory.get("url") or not memory.get("api_key"):
+        raise ValueError(f"voice user {user_id} has no valid memory profile")
+    return {**profile, "user_id": user_id}
+
+
+def user_id_from_room(room_name: str) -> str:
+    """Extract the trusted user slug minted into a server-generated room name."""
+    match = re.fullmatch(r"voice-([a-z0-9][a-z0-9_-]{1,30})-[0-9a-f]{8}", room_name)
+    if not match:
+        raise ValueError(f"invalid voice room name: {room_name}")
+    return match.group(1)
+
 
 SYSTEM_PROMPT = textwrap.dedent("""\
-    You are Hope, a warm, conversational voice assistant. You are talking TO
-    someone, not writing for them to read. Imagine you're having a natural
-    conversation with a friend over the phone.
+    You are Hope, a warm, conversational voice assistant speaking with
+    {user_display_name}. You are talking TO someone, not writing for them to
+    read. Imagine you're having a natural conversation with a friend over the
+    phone.
 
     # How you speak
     - Keep every response to one or three sentences. That's it.
@@ -81,21 +106,21 @@ SYSTEM_PROMPT = textwrap.dedent("""\
     - Never read back URLs, file paths, or technical identifiers.
 
     # Memory (CRITICAL — ALWAYS call these tools directly, never dispatch)
-    When the user asks about personal information, past conversations, or
-    anything you might have stored, you MUST call recall FIRST before
-    answering. Do NOT say "I don't have that info" without calling recall.
-    When you call recall, ALWAYS include session_id="hope" to search the
-    session cache.
-    When the user tells you something to remember or shares a preference,
-    you MUST call remember with session_id="hope". Always include
-    session_id="hope" in every remember call — this makes it instant.
-    - User says "what's my name?" → CALL recall with query="my name", session_id="hope"
-    - User says "do you remember who I am?" → CALL recall with query="user identity name", session_id="hope"
-    - User says "remember my coffee order is oat milk latte" → CALL remember
-      with data="User's coffee order is an oat milk latte", session_id="hope"
-    - User says "forget that I like blue" → CALL forget to remove it
-    After calling remember, confirm briefly ("Got it, I'll remember that").
-    If recall returns nothing, then say you don't have that info yet.
+    Your memory tools are private to the current user. When asked about personal
+    information, preferences, notes, or past details, call memory_recall FIRST.
+    Never claim you lack the information before searching memory.
+    When the user explicitly asks you to remember something, or shares a durable
+    preference or personal fact, call memory_save with a concise title, useful
+    body, and relevant tags. Do not save every utterance.
+    - "What's my name?" → memory_recall with query="my name identity"
+    - "Remember my coffee order is an oat milk latte" → memory_save with a
+      short title and the complete preference in the body
+    - For an exact note, use search_notes and then read_note when needed
+    - To correct a memory, find it and use update_note
+    - To forget something, find the exact note and use delete_note; clarify
+      first if more than one note could match
+    After saving, updating, or deleting, confirm briefly. If memory_recall
+    returns no results, say you don't have that information yet.
 
     # Weather & Time
     You can check the weather for any location. Use get_weather when the user
@@ -110,8 +135,8 @@ SYSTEM_PROMPT = textwrap.dedent("""\
     This includes: looking up news, researching topics, checking current
     events, prices, sports scores, writing something, summarizing, comparing
     options, planning, or anything that takes more than a couple seconds to
-    think through. Memory recall/remember is NEVER a dispatch task —
-    call those tools directly.
+    think through. Memory tool calls are NEVER dispatch tasks — call those
+    tools directly.
     Examples:
     - User says "what's the latest news?" → CALL dispatch_task with
       description="Find the top 5 international news headlines today"
@@ -361,13 +386,8 @@ class GemmaLLM(openai.LLM):
 
 # ── MCP toolsets (web access + any extra configured servers) ────────────────
 
-def build_mcp_toolsets() -> list[mcp.MCPToolset]:
-    """Build the list of MCP toolsets to attach to the agent.
-
-    Always includes the local web-access server (Firecrawl-backed search/scrape)
-    when WEB_MCP_ENABLED is true. Additional servers can be configured via the
-    EXTRA_MCP_SERVERS env var (JSON list of {url, transport} objects).
-    """
+def build_mcp_toolsets(user_profile: dict) -> list[mcp.MCPToolset]:
+    """Build isolated MCP toolsets for one authenticated voice user."""
     toolsets: list[mcp.MCPToolset] = []
 
     python_bin = os.path.join(os.path.dirname(os.path.abspath(__file__)), ".venv", "bin", "python")
@@ -385,19 +405,38 @@ def build_mcp_toolsets() -> list[mcp.MCPToolset]:
     )
     logger.info("Weather MCP toolset enabled (wttr.in)")
 
-    cognee_url = os.environ.get("COGNEE_MCP_URL", "http://192.168.86.2:8003/mcp")
+    memory = user_profile["memory"]
+    memory_url = memory["url"]
+    allowed_memory_tools = memory.get(
+        "allowed_tools",
+        [
+            "create_note",
+            "read_note",
+            "update_note",
+            "delete_note",
+            "list_notes",
+            "search_notes",
+            "memory_save",
+            "memory_recall",
+        ],
+    )
     toolsets.append(
         mcp.MCPToolset(
             id="memory",
             mcp_server=mcp.MCPServerHTTP(
-                url=cognee_url,
+                url=memory_url,
                 transport_type="streamable_http",
-                allowed_tools=["remember", "recall", "forget"],
+                allowed_tools=allowed_memory_tools,
+                headers={"Authorization": f"Bearer {memory['api_key']}"},
                 client_session_timeout_seconds=60,
             ),
         )
     )
-    logger.info("Memory MCP toolset enabled (Cognee at %s)", cognee_url)
+    logger.info(
+        "Memory MCP enabled for user %s at %s",
+        user_profile["user_id"],
+        memory_url,
+    )
 
     skills_mcp_script = os.path.join(os.path.dirname(os.path.abspath(__file__)), "skills_mcp.py")
     toolsets.append(
@@ -453,7 +492,9 @@ def build_mcp_toolsets() -> list[mcp.MCPToolset]:
 class VoiceAssistant(Agent):
     """The conversational agent. LLM is the brain; STT/TTS are senses."""
 
-    def __init__(self) -> None:
+    def __init__(self, user_profile: dict) -> None:
+        display_name = user_profile.get("display_name") or user_profile["user_id"]
+        instructions = SYSTEM_PROMPT.replace("{user_display_name}", display_name)
         super().__init__(
             llm=GemmaLLM(
                 model=GEMMA_MODEL,
@@ -462,8 +503,8 @@ class VoiceAssistant(Agent):
                 max_completion_tokens=1000,
                 extra_body={"chat_template_kwargs": {"enable_thinking": False}},
             ),
-            instructions=SYSTEM_PROMPT,
-            tools=build_mcp_toolsets(),
+            instructions=instructions,
+            tools=build_mcp_toolsets(user_profile),
         )
 
 
@@ -474,6 +515,13 @@ server = AgentServer()
 @server.rtc_session(agent_name="voice-assistant")
 async def handle_job(ctx: JobContext) -> None:
     logger.info("Job started for room %s", ctx.room.name)
+    try:
+        user_id = user_id_from_room(ctx.room.name)
+        user_profile = load_user_profile(user_id)
+    except (OSError, ValueError, json.JSONDecodeError) as exc:
+        logger.error("Rejecting voice job: %s", exc)
+        return
+    logger.info("Resolved voice user %s", user_id)
 
     # Azure STT — streaming, reads AZURE_SPEECH_KEY / AZURE_SPEECH_REGION from env
     stt = azure.STT(
@@ -589,7 +637,7 @@ async def handle_job(ctx: JobContext) -> None:
     asyncio.create_task(_poll_tasks())
 
     await session.start(
-        agent=VoiceAssistant(),
+        agent=VoiceAssistant(user_profile),
         room=ctx.room,
         room_options=room_io.RoomOptions(
             # Keep the agent in the room when a participant leaves; it must