Failure Modes of Existing Memory Methods

1 MemGPT · flat FIFO queue pages of every topic share one queue, so topics mix as the dialogue grows OS-style paging, but no notion of a topic in → one queue · topics interleaved oldest out, regardless of topic 2 MemoryBank · decay alone a forgetting curve decides what fades, but nothing organizes or updates the rest worst scores of all compared methods strength only decays · no structure, no update 3 TiM · single-stage hash retrieval one hash lookup returns similar thoughts, but cannot preserve cross-topic dependencies saving thoughts beats raw turns, retrieval is flat topic A park topic B diet goal ✗ the link between topics is lost one lookup, one bucket at a time 4 A-Mem · heavy link generation every note needs multi-step LLM calls to link, which inflates latency and accumulates errors 13.0 LLM calls per response (Table 3) LLM call × 13 per response each step can inject an error rich graph, slow and fragile to build (MemoryOS: 4.9 calls) each method works on a single dimension in isolation · storage structure, retrieval mechanism, or update strategy no unified system that stores, updates, retrieves and generates together
07 / 25