agentic-memory-support-demo

A small chatbot demo that shows why agentic-memory matters.

This repo pairs a support-style chatbot with a deterministic benchmark harness. The benchmark compares the exact same bot in two modes:

• baseline: no long-term memory
• memory_enabled: semantic, episodic, and procedural memory via the agentic-memory SDK

The point is not "chatbot vibes". The point is to show measurable behavior changes:

• remembering user profile facts across turns
• not asking users to repeat troubleshooting steps
• retrieving the right procedure for the current issue

What This Demonstrates

Most chatbot demos look better because the prompt got longer or the examples got hand-picked.

This repo tries to isolate the contribution of memory itself:

• the same deterministic LLM is used in both benchmark modes
• the only difference is whether the bot can write and retrieve memories
• scoring is exact substring matching, not an LLM judge

Deterministic Benchmark

The benchmark includes 12 synthetic support scenarios:

• 4 current-turn control cases
• 4 cross-turn profile recall cases
• 2 troubleshooting continuity cases
• 2 procedure retrieval cases

Expected benchmark result:

Metric	Baseline	Memory Enabled
Overall accuracy	33.3%	100.0%
Current-turn controls	100.0%	100.0%
Profile recall	0.0%	100.0%
Troubleshooting continuity	0.0%	100.0%
Procedure retrieval	0.0%	100.0%

Quickstart

1. Create a virtual environment

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e .
export GEMINI_API_KEY=your_key_here

2. Run the benchmark

memory-support-benchmark

3. Run the Streamlit demo

streamlit run streamlit_app.py

If GEMINI_API_KEY is set, the app uses Gemini both for chatbot generation and for agentic-memory embeddings. If not, it falls back to a deterministic responder so the demo still works offline.

Why The Demo Is Structured This Way

This repo is intentionally opinionated:

• semantic memory stores durable user facts like plan, timezone, and preferences
• episodic memory stores troubleshooting history and prior conversation events
• procedural memory stores reusable support runbooks

That makes the value proposition legible in one sentence:

the bot remembers who the user is, what already happened, and what usually works

Project Layout

agentic-memory-support-demo/
├── evals/
│   └── run_benchmark.py
├── src/
│   └── agentic_memory_support_demo/
│       ├── __init__.py
│       ├── benchmark.py
│       ├── chatbot.py
│       ├── deterministic.py
│       └── llm.py
├── tests/
│   └── test_benchmark.py
├── pyproject.toml
└── streamlit_app.py