When Does Graph Memory Help Beyond Verbalized Relations? Isolating Representation from Structure in LLM-Agent Memory
Independent research — not committed as a thesis topic
- Question
- Graph memory is widely believed to help LLM agents remember better. But when graph memory is compared against flat memory, existing work changes TWO things at once: structure (the ability to traverse relations) and representation (information rewritten as typed relational sentences — unavoidable, because an LLM can only read text). So which of the two does the observed benefit belong to?
- Method
- Verbalize-control: three token-budget-matched conditions drawing from the same candidate edge pool and differing only in the selector — graph traversal, cosine truncated to the budget, and dump-everything. Plus a degree-preserving shuffled-edge control to separate “these particular edges” from “merely having a graph”. The design is pre-registered, evaluated on an external non-circular benchmark, and specifies the negative branch in advance (TOST equivalence testing).
Status, stated plainly: this is something I work on outside of client work, not a committed thesis topic. A 30-question feasibility probe has been run; the quality gate fired and returned “inconclusive” rather than a result. No publishable numbers.
