CUHK Paper: Agentic Markdown Memory Is 'Memos, Not Memory'
Researchers at the Chinese University of Hong Kong (Hangzhou) published a formal paper (April 30, 2026) drawing a sharp line between context-based retrieval systems (skill MDs, RAG stores, scratchpads) and parametric weight-based learning. Their compositional sample-complexity theorem shows that retrieval requires explicit coverage of all concept combinations (combinatorial blow-up), while weight updates generalize from O(N) examples. The "frozen novice" problem: without weight updates across sessions, the model never becomes more expert. They also flag long-term retrieval memory as a prompt-injection persistence vector.
Why It Matters
The "memos vs. memory" framing gives practitioners a clear vocabulary for the architectural ceiling of purely retrieval-based agent memory systems — and identifies continual fine-tuning of small open models as the only path to genuine accumulation.