Memory That Doesn't Expire: Agentic Long-Term Memory for Conversational Agents on a Single Consumer GPU

Conversational agents forget: once a dialogue outgrows the context window, a model must either be reminded through retrieval or have its history truncated. Recent memory systems address this with extraction and retrieval, but they assume a large, often API-hosted model as the memory manager, an assumption that rules out private, local deployment. We ask whether a small local model, on a single consumer GPU, can manage long-term memory well, and how the answer depends on model size. We evaluate four memory strategies (no memory, full-history stuffing, raw-turn retrieval, and an agentic protocol that extracts typed, dated, deduplicated facts) on LoCoMo (1,986 questions over 10 long multi-session conversations), scored by an independent LLM judge. Full-context stuffing wins aggregate accuracy when the answer still lies inside the window, but its advantage is an artifact of conversation length: split by whether a question's evidence survives truncation, stuffing collapses by 76% beyond the window while agentic memory stays flat, outperforming stuffing 2.0x on the 39% of questions that already exceed it. Agentic memory further gives the best temporal reasoning of any strategy (via absolute-date grounding at write time) using ~67x less context. Varying the extractor from 3B to 14B, we find quality scales with size and that temporal grounding is the first capability to break under compression, setting a minimum viable memory-manager size of ~7-14B, comfortably within a 12 GB GPU. A hand-verified audit of 200 stored memories finds 199 faithful to their source turn, with no fabrications. We release the system, the benchmark harness, and the raw judged outputs behind every reported number. Code: https://github.com/MasHasNoLife/yuna-aiPaper source: https://github.com/MasHasNoLife/memory-that-doesnt-expireRaw run outputs: https://doi.org/10.5281/zenodo.21702651