mem0 conversation manager demo

• Python 3.13 venv in .venv
• Implements:
  1. Summarize every 6 user–assistant turns (~200 tokens) into a stripe
  2. Each new query loads 6 most relevant memories from mem0ai
  3. Total context = system prompt + long-term memory + stripes + compressed history + running history + latest query
  4. If context > 12k tokens, compress earliest ~6k tokens into ~500-token summary
  5. Long-term memory writes happen through the registered tool when required

Setup

pwsh python3.13 -m venv .venv .\.venv\Scripts\pip install --upgrade pip .\.venv\Scripts\pip install -r requirements.txt

Set environment variables before running:

• KIMI_API_KEY (or use provided NEXT_PUBLIC_KIMI_API_KEY)
• KIMI_BASE_URL (defaults to https://api.moonshot.cn/v1)
• MEM0_API_KEY (optional if using hosted mem0)
• PRELOADED_STATE_FILE (optional) – path to preprocessed stripes/compressed history JSON

Ingest historical conversations

If you exported all chats into conversations.json, preprocess them into a compact state:

pwsh .\.venv\Scripts\python ingest_conversations.py conversations.json preloaded_state.json

This streams the large JSON with ijson, replays the dialogue with lightweight summarisation/compression, and stores:

• stripes: ~200-token summaries of every 6-turn window
• compressed: previously-compressed history chunks (~500 tokens each)
• history: the most recent uncompressed messages

You can limit ingestion for testing:

pwsh .\.venv\Scripts\python ingest_conversations.py conversations.json preloaded_state.json --limit 10 --dry-run

Finally set PRELOADED_STATE_FILE=preloaded_state.json before running to hydrate the manager with stripes and history.

Run demos

Baseline CLI (JSON output):

pwsh .\.venv\Scripts\python -m src.main

Rich console UI（图形化展示 token & 上下文来源）:

pwsh .\.venv\Scripts\python -m src.ui_console

Type /ltm during the session to force a long-term memory write.