Short-Term Memory (SU) Architecture

Efficient Human-Like Memory Management

Why Short-Term Memory Matters

Loading entire conversation histories is extremely inefficient and becomes unsustainable as threads grow long. The SU system mimics how humans naturally remember: recent events stay easily accessible, while older ones are recalled only when specifically needed.

How SU Short-Term Memory Works

• SU Tag Layer: Recent Memory Units are tagged as “SU” (Short-Term Use) and kept readily available.
• FIFO Management: Oldest SU-tagged memories automatically drop off on a customizable schedule (default example: 2 weeks). This keeps the active context lean and fast.
• “Do You Remember” Invocation: When older context is needed, the PI can be asked “Do you remember…?” to pull specific MUs from long-term storage by tag or index. This is fast and targeted.
• “Keep in Mind” Function: Important items that should persist beyond the normal SU window can be explicitly flagged. These stay active even after the normal FIFO drop.

Handling High-Addendum MUs

Memory Units that accumulate many addendums (frequent updates or deep refinement) are automatically recognized as high-value. These can be indexed with higher priority or automatically moved into “Keep in Mind” status so they remain easily accessible regardless of age.

Benefits of This Approach

• Much faster startup and response times (far fewer tokens loaded initially)
• More natural, human-like conversation flow
• Prevents context bloat while preserving access to important history
• Customizable per user or node (duration, importance thresholds, etc.)
• Better overall token efficiency and cost management

Best Practices

• Use clear “Do you remember” or “Keep in mind” phrasing when needed
• Regularly review and tag truly important long-term MUs
• Adjust SU duration based on usage patterns (heavier sessions may need shorter windows)
• Trust the system — most recent context is usually what matters most

Return to Main Mark V Page