Auto Save Memory Plugin

Automatically extracts and saves strategic insights from conversations into long-term project memory — without any manual memory_save calls.

How It Works

1. Hooks into monologue_end — fires after each agent response turn
2. Runs in background — zero impact on chat response time or latency
3. Uses utility LLM — the cheap/fast model (e.g. gpt-4o-mini), not your main chat model
4. Saves to Memory.Area.MAIN — recalled alongside solutions in future turns
5. Project-aware — tags each saved memory with the active project name for filtering

What Gets Saved

The utility LLM extracts only durable, self-contained insights:

Trivial exchanges, generic statements, and filler conversation are ignored.

Configuration

Edit default_config.yaml or override in your project/agent config:

auto_save_enabled: true        # Master on/off switch
auto_save_interval: 1          # Run every N turns (1 = every turn)
auto_save_min_history_len: 200 # Skip if history is too short
auto_save_memory_area: main    # Memory area: 'main' or 'fragments'
auto_save_history_len: 8000    # Characters of history to analyze

Cost

• Main chat LLM cost: Zero impact
• Utility LLM cost: ~$0.001 per turn with gpt-4o-mini or gemini-flash
• If nothing to save: The LLM call still happens but returns [] — negligible cost
• To reduce cost: set auto_save_interval: 3 to run every 3rd turn instead

Differences from Built-in Memory

Both run independently and complement each other.

Viewing Saved Insights

You can query saved insights at any time:

memory_load query="AgriKonnect strategy decisions" filter="source=='auto_save_memory'"

Disabling

Set auto_save_enabled: false in the plugin config, or disable the plugin entirely from the Agent Zero plugin manager.