Retrieval is only one use of memory
Enhancing Local and Global Structure in Compressed AI Agent Memory argues that compression should be evaluated beyond the nearest neighbors returned by a search. Agents also group, compare and delete memories, operations that depend on relationships across a larger portion of the collection.
The paper measures local fidelity at ten neighbors and global fidelity across a wider band of the memory. This distinguishes a representation that retrieves nearby passages well from one that preserves their broader organization.
A strong local score can conceal a structural change
At 96 bytes per passage, the study reports a rescored binary representation with local fidelity of 0.97 that moves 10.1% of passages to another theme. A PCA-based eight-bit representation has lower local fidelity in the comparison but much less theme movement.
These figures belong to the evaluated memory and encoding conditions. They do not establish that binary compression is always unsuitable or that a global score alone determines an agent’s practical performance.
What locality-aware quantization changes
The proposed LAQ method modifies an OPQ index’s codebook objective while retaining its byte budget, rotation and search arrangement. The authors report improved global fidelity in all 52 held-out runs and retained evidence recall in their tests.
Franklin’s analysis is that a memory system’s acceptance test should reflect the operations it actually performs. If it only retrieves short passages, local recall may dominate. If it consolidates themes or removes a client’s information, broader structure and the correctness of those actions need attention too.
Readers should examine the label-free fidelity definitions, datasets and downstream checks before adopting the objective. The preprint provides a mathematical argument and experimental comparisons, rather than proof that a single compression setting is optimal for every agent memory workload.
Comments