Information Flow Understanding in Oblivion
Running example demonstrating the Oblivion pipeline across a 34-turn conversation covering seven topics (travel, cooking, work/technology, health/fitness, reading, astrophotography, mushroom foraging) with repeated questions to show Ebbinghaus-style memory decay and reinforcement spikes. Each turn is one back-and-forth exchange (one human message + one agent response). Step through each turn to see module states, buffer contents, and all logged metrics.
Conversation
Loading conversation data…
Module State
Loading module state…
Metrics Overview
Aggregated visualizations across all 34 interactions. Data is pre-generated from a live Oblivion session with full metric logging enabled.
Aggregated Metrics Summary
▼Loading aggregated metrics…
Latency Breakdown
▼Stacked bar showing per-module processing time. Each segment represents Recognizer, Decayer, Activator, and Response generation latency for that turn.
Token Usage
▼Prompt vs completion tokens consumed per interaction, aggregated across all modules.
Buffer Size Over Time
▼Total memories in the working memory buffer over time. Growth indicates new memory extraction; drops indicate purging of low-decay memories.
Activator Trigger Events
▼Number of times the Activator was invoked per turn. activator_trigger_count counts actual Activator invocations, while thinking_iterations_used counts total Decayer evaluation rounds (the Decayer may decide the buffer is sufficient without triggering the Activator).
Decay Curves — Semantic Memories
▼Reinforced semantic memory decay scores over turns. Decay follows the Ebbinghaus curve R = e-t/(S·T); spikes indicate reinforcement when a memory is used in a response.
Decay Curves — Episodic Memories
▼Reinforced episodic memory decay scores over turns. Only episodic entries with unique content (different from semantic) are shown. Each legend shows the memory start…end with word count.