Agent Memory Architectures

Memory is the harness component that turns a sequence of calls into an agent with continuity. The organising frame is write-manage-read — what gets committed, how it is curated and evolved, and how it is retrieved — spanning five mechanism families under a three-dimensional taxonomy.

The manage stage is the one most often skipped and the one that decides whether memory helps. A store that only writes and reads accumulates until retrieval degrades; the curation step is what keeps it useful. A-MEM is the representative design: Zettelkasten-inspired note construction with dynamic linking and memory evolution, so stored notes are restructured as new material arrives rather than merely appended.

Memory is also where the security and cost stories meet the architecture. Persistence is exactly what makes memory poisoning compound (see the exploitation surface) — a poisoned memory keeps paying the attacker across later runs. And every token of retrieved context is a token paid for on every subsequent call, which is where memory design lands on inference-economics.

Key Claims

• Write-manage-read taxonomy across five mechanism families organises agent memory designs. Evidence: moderate (paper) (Agent Memory Mechanisms Survey)
• A-MEM: Zettelkasten-inspired notes with dynamic linking and memory evolution — memory is restructured, not merely appended. Evidence: moderate (paper) (A-MEM)