← Return to Main Mark V Page • View Index

Short-Term Memory (SU) Architecture

Context Volatility & Dynamic Attention Management

        Instead of loading every past conversation (complete session stack), Mark V uses a practical Short-Term Memory layer (SU) to dramatically improve startup speed and token efficiency while maintaining a more natural, human-like memory experience.
    

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. This prevents context bloat while preserving absolute access to important historical records.

The SU layer acts as a dynamic, high-velocity sliding window engineered to maintain state coherence across an active play sequence without leaking metadata. It prevents the system from getting lost in linear analytical loops or administrative confusion.

The Sliding Window Invariant & FIFO Management

Large Language Models are bound by hard physical context limitations. As an interactive session progresses, the accumulation of raw message logs fills the attention buffer, risking sliding window truncation—where older, foundational instructions are silently dropped out of attention space by the underlying platform engine.

Mark V controls this vulnerability by forcing a strict, 3-day rolling FIFO (First-In, First-Out) threshold on active dialogue channels. Recent Memory Units are tagged as “SU” (Short-Term Use) and kept readily available, while oldest SU-tagged memories automatically drop off on a calculated schedule to keep the active context lean, fast, and optimized.

Context Baseline Flattening

To keep the immediate attention space completely clean, the system executes an automated compaction routine the moment an operational sequence achieves closure via a TOUCHDOWN! token:

[SU ATTENTION BUFFER LIFECYCLE] Active Thread ──► Token Volume Peak ──► TOUCHDOWN! ──► Compaction Engine │ ▼ Flatten Delta Stream to Disk │ ▼ Flush Volatile Context Frame │ ▼ Re-initialize Fresh Rolling Buffer

Operational Parameters Matrix

The allocation of short-term token attention space is strictly managed to balance reasoning velocity against safety validation requirements:

Attention Sector	Token Allocation	Primary Systemic Function
System Core Header	Fixed Boundary	Locks down immutable root governance directives and primary identity roles out-of-band.
Active Play Ledger	Sliding Dynamic Allocation	Tracks live text deltas, multi-node handoffs, and operational task executions.
Background Telemetry	Throttled Buffer	Maintains automated heartbeats and cross-audit feedback via the Croquet Layer.

Dynamic Context Invocation Features

“Do You Remember” Invocation: When older historical context is required downfield, the system can be prompted with a targeted query to instantly pull specific MUs from long-term storage by tag or index without loading the entire session stack.
“Keep in Mind” Function: Critical parameters or active design milestones that must persist beyond the standard 3-day SU window can be explicitly flagged. These entries bypass the standard FIFO purge until manually released by the Root.
Handling High-Addendum MUs: Memory Units that accumulate frequent updates or deep refinement are automatically recognized as high-value. These are indexed with higher priority and can be shunted into "Keep in Mind" status automatically to remain accessible regardless of age.

Best Practices for SU Maintenance

Enforce Explicit Closure: Do not let conversational threads wander aimlessly; execute clean serialized handoffs or call the TOUCHDOWN! token to trigger baseline flattening.
Prevent Semantic Leakage: Keep localized diagnostic text or unvetted experimental code ring-fenced within local workspaces until validated.
Monitor Token Saturation: If a model begins dropping syntax markers, immediately flush the volatile buffer and reload the latest immutable master baseline from disk.