Agent Memory Costs: Characterizing Write-Time Overhead in Long-Horizon Workloads

Omri, Gan, Broveak, Geens, He, Pentland, Verhelst, Weissman, Tambe · 2026 · "Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads" arXiv 2606.06448

Seven-stage memory pipeline

Click a stage. Construction is the write path, the hidden bill.

Four construction styles

One chat turn: "I moved to Lisbon last month."

Why not paste everything?

Three failure modes of full-history long context.

Ten systems on a write-time LLM gradientIllustrative ordering

Order follows the episode's list. Colour encodes rank, not a measured value. Hover or tap a row.

Benchmark setup

LongMemEval_S_* inside MemoryAgentBench, reported figures.

Results explorer

Pick a metric. All values are numbers stated in the episode.

Tail latency: p95 / medianReported

Colour by ratio (blue low, red high).

Phase-tagged GPU timelineIllustrative

Schematic of the call pattern, not measured traces. Toggle the loop style.

Long-read, short-writeReported

Co-tenancy: shared backend throughputReported

Ten independent builds on one backend, relative to single-build throughput.

Freshness: block or go staleIllustrative

Single-worker queue sketch. Slide the gap between sessions and per-session construction time. The paper reports per-session construction from ~1 ms to ~10 s.

Amortization anchorEpisode arithmetic

Mem0 one-sample build ≈ 1,050 kJ (from the episode) spread over N queries. The paper does not sweep N; this is only 1050 / N. Per-query saving versus long context is not in the cited table.

How firm is each claim?Episode's assessment

Hover a claim.

References

  1. Omri et al., Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads, 2026 — arXiv:2606.06448
  2. Packer et al., MemGPT: Towards LLMs as Operating Systems, 2023 — scholar
  3. Chhikara et al., Mem0, 2025 — scholar
  4. Maharana et al., LoCoMo, 2024 — scholar
  5. Lin et al., Sleep-time Compute, 2025 — scholar
  6. Rasmussen et al., Zep, 2025 — scholar
  7. Cheng et al., LMCache, 2025 — scholar
  8. Qin et al., Mooncake, 2024 — scholar
  9. Agrawal et al., Sarathi-Serve, 2024 — scholar
  10. LightMem and other efficiency-oriented memory systems, 2025 — scholar
  11. Eyuboglu et al., Cartridges, 2025 — scholar
  12. Kwon et al., PagedAttention (vLLM), 2023 — scholar
  13. Verma et al., Borg, 2015 — scholar