Can three axes replace the short-term long-term memory split?

The agent-memory literature has fragmented faster than the field can taxonomize. "Memory in the Age of AI Agents" (2512.13564) argues the older long/short-term split was the source of confusion: it conflates where memory lives, what role it plays, and how it changes — three independent questions that need three separate taxonomies.

Forms answer where memory physically lives: token-level (memory as text in the prompt or external store retrieved into the prompt), parametric (memory as model weights, including fine-tuned adapters and PKM-style key-value layers), or latent (memory as continuous hidden states preserved across turns or sessions). A single agent can use all three simultaneously.

Functions answer what role memory plays: factual (a fact about the world the agent uses as a lookup), experiential (a trajectory of past actions and outcomes the agent learns from), or working (the active scratchpad holding the current goal, plan, and intermediate results). The same physical store can serve different functions at different times.

Dynamics answer how memory changes: formation (selecting which artifacts from a step become memory candidates), evolution (integrating candidates into the existing store via consolidation, conflict resolution, or pruning), retrieval (constructing task-aware queries to surface relevant content). Crucially, short-term and long-term phenomena emerge from temporal patterns of these operators, not from architecturally separate modules.

The reframing is consequential. The old taxonomy invited claims like "this system has long-term memory because it has a vector database" — implementation by furniture, not behavior. The new taxonomy forces specificity: this system stores experiential memory in token form with eager-formation, periodic-evolution, and similarity-based-retrieval dynamics. Comparing systems becomes possible because they describe themselves along comparable dimensions rather than each inventing its own vocabulary.

The survey also clarifies the scope boundary with traditional LLM memory work. KV-cache management, long-context extensions, and architectural changes to retention during inference remain LLM-internal memory — not agent memory — because they address sequence processing rather than goal-directed behavior persisting across tasks.

Paper: Memory in the Age of AI Agents: A Survey