Two real-environment observations from the memory-system e2e (design: docs-dev/copilot/robust-browser-agent-memory-system-design.md, section 16):
-
L1 knowledge never appears in the recall section — repeated example.com tasks only produce [L0] hits + user prefs; the PEM layer classifies the intent as OTHER (low confidence), so [L1] stays filtered. Needs a multi-round same-site run to observe whether confidence grows and [L1] shows up; if not, revisit intent classification / URL-pattern matching for the recall query.
-
Short tasks miss cli-prompt tracing — a 14s task (memory_search + memory_read + taskComplete) produced events in the memory log but NO cli-prompt/*.request.json under the agent trace dir. The CliLoopTracer prompt-dump timing loses very short runs, hurting auditability.
Both are observation/verification items requiring a real LLM + running backend; no code change planned until observed.
Two real-environment observations from the memory-system e2e (design: docs-dev/copilot/robust-browser-agent-memory-system-design.md, section 16):
L1 knowledge never appears in the recall section — repeated example.com tasks only produce [L0] hits + user prefs; the PEM layer classifies the intent as OTHER (low confidence), so [L1] stays filtered. Needs a multi-round same-site run to observe whether confidence grows and [L1] shows up; if not, revisit intent classification / URL-pattern matching for the recall query.
Short tasks miss cli-prompt tracing — a 14s task (memory_search + memory_read + taskComplete) produced events in the memory log but NO cli-prompt/*.request.json under the agent trace dir. The CliLoopTracer prompt-dump timing loses very short runs, hurting auditability.
Both are observation/verification items requiring a real LLM + running backend; no code change planned until observed.