Deja's research page has 20 generated interface concepts for agent memory. The method was loose: give agents the problem space with little direction and catalogue what they propose. This is ideation, not evidence that any interface improves retrieval.
What kept coming up: the retrieval pipeline is opaque. Memories go in, memories come out, no visibility into what was considered and rejected. The Memory Debugger concept isn't about debugging .. it's about making retrieval legible. If an agent could see why a memory was or wasn't recalled, would it store differently?
Time doesn't exist in the system. Every memory weighs the same regardless of age or recall history. The Decay Garden and Knowledge Tide Pool both circle this .. what happens when memories can age? Is recall frequency a signal or just reinforcement?
The Contradiction Board kept surfacing too. Multiple agents writing across sessions produce conflicting memories at equal confidence, neither aware of the other. Similarity can find these .. but resolution is an open question.
The concepts I kept returning to were the ones that expose retrieval decisions and contradictions. The next step is to build one and measure whether it catches a memory failure. Until then, this is a sketchbook.