alexa-memory-mcp

A Model Context Protocol (MCP) server that gives Alexa+ a persistent memory. Voice assistants forget everything between sessions: you tell them your dog's name on Monday and by Friday they have no idea. This project wraps a knowledge-graph memory engine (facts stored as subject/relation/object triples in a graph database, plus the raw memories they came from) behind four MCP tools, so an Alexa+ integration can learn facts once, recall them in any later session, forget on request, and report what it knows.

Why Alexa+ needs this

Alexa+ integrations are built on the open MCP standard, but the standard only carries messages: there is no durable memory layer. Without one, every conversation starts from zero and personalization is impossible. alexa-memory-mcp plugs that gap with a persistent graph memory: tell the assistant something once ("my dog's name is Bruno"), and it stores a structured triple (my dog / has_name / Bruno) linked to the original statement. Later sessions ask questions against the graph, retrieve exactly the relevant facts, and answer from them instead of guessing. Forgetting is first-class too: memory_forget wipes an entity and everything attached to it, which is what privacy-conscious voice users expect.

Related MCP server: Memory Server MCP

Architecture

 demo page (index.html + app.js)
        |
        |  HTTP (demo) / MCP over Streamable HTTP (Alexa+ clients)
        v
 FastMCP server (server.py)
        |
        v
 MemoryAgent (memgraph/agent.py) --- tell / ask / forget / stats
        |                                      |
        v                                      v
 InMemoryGraphStore or FalkorDB        LLM provider
 (zero-setup for demos /              (Gemini free tier, OpenAI-compatible,
  real graph DB in prod)               or FakeLLM offline)

The demo page talks to the same MemoryAgent the MCP tools use, so what you see in the browser is exactly what Alexa+ would get.

Setup and run

Requirements: Python 3.10+.

cd alexa-memory-mcp
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Environment variables:

Run offline (no keys, no database):

LLM_BACKEND=fake python server.py

Then open http://127.0.0.1:8765/ for the demo page: tell it a fact, ask about it later, watch the triple land in the graph.

For production memory, point the server at FalkorDB (FALKORDB_HOST / FALKORDB_PORT / FALKORDB_PASSWORD) and set a real LLM_BACKEND with the matching API key.

Run the tests

python -m venv .venv && source .venv/bin/activate
pip install pytest
python -m pytest tests/ -v

Tools

The four MCP tools exposed by server.py:

Example session:

tell("my dog's name is Bruno")     -> [Triple(subject="my dog", relation="has_name", object="Bruno")]
ask("what's my dog's name?")       -> "Bruno"
stats()                            -> {"triples": 1, "memories": 1, "entities": 2}
forget("my dog")                   -> True

Built during the hackathon window

Build log for the hackathon period (all work landed Oct 17, 2026, ahead of the Oct 23 deadline).

• 2026-10-17: Scaffolded the project. Built the MCP server (server.py, FastMCP, spec 2025-11-25, Streamable HTTP, four tools). Vendored the memgraph knowledge-graph engine. Built the simulated Alexa+ web demo (demo/). Wrote the pytest suite (5 tests). Fixed: Gemini default model (gemini-2.0-flash 404s for new keys, moved to gemini-3.8-flash); recall fallback for coreference resolution ("my cat" -> "Whiskers"); 429 backoff retries; extract prompt has_name rule; answer prompt name-resolution guidance. Recorded the 66-second demo video. Published the repo (public, MIT). Submitted on Devpost (Alexa+ track + Open Source mini challenge).