What's this about?
Memory is becoming core infrastructure for long-horizon agents, but existing evaluations give little guidance on which memory *substrate* to use — the underlying medium in which memory is represented and stored. This paper (12 upvotes) evaluates substrates holistically rather than comparing whole memory systems.
What exactly is a memory "substrate"?
It's what the memory physically is. Options include a vector database storing embeddings you search by similarity; a knowledge graph storing entities and relationships; plain text files the agent reads and rewrites; a relational database with structured records; or simply a long context window. Each has different properties for what's easy to store, easy to retrieve, easy to update, and easy for a human to inspect. Most comparisons pit complete systems against each other, which confounds the substrate with everything built on top — this isolates the layer underneath.
Why is this worth noticing?
- Memory is the acknowledged missing capability — it's what Rich Sutton's new lab targets, what Chain-of-Experience measures, and what MindMemOS and Cumora build around — yet the basic engineering question of what to store it in is unsettled.
- The Pi team's position from last week is a live entry in this debate: "code is the truth," meaning for a codebase the substrate should be the repository itself rather than an index over it.