agentic-memory

A learning-focused CLI chatbot that demonstrates every major AI-agent memory type: short-term, episodic, semantic (with contradiction detection + user confirmation), procedural, plus a query router that pulls the right memory into every prompt. Built incrementally as vertical-slice stories S01–S10. No automated tests — manual verification is the acceptance criterion for every story.

Setup

The first LLM/HF model call downloads the sentence-transformer and NLI weights (~500 MB). Pre-warm them if desired:

Automated 5-Session Walkthrough

Between sessions, a BEFORE/AFTER diff table prints memory counts and lists pipeline stages that fired. A WALKTHROUGH event is emitted to logs/walkthrough.jsonl around each session for later analysis.

Manual Walkthrough (5 Sessions)

Designed for a live audience demo. Each session has: (1) reset/setup, (2) exact user inputs, (3) what the audience should watch for on-screen while typing, (4) verification queries to run after /exit, and (5) the memory concept it demonstrates. Keep two terminals open — one running main.py, another for sqlite3 and log_view.py.

Preflight — Start clean

Now start the chatbot in terminal A:

Note the session line printed at startup — that's the session id for log_view.py.

Session 1 — First Contact • short-term buffer + episodic write + fact extraction

Type into terminal A (one line at a time):

Watch on screen (terminal A) as you type: