---
source: https://iinventedlight.com/library/notes/ao-memory-long-term-benchmarks.md
title: "How well do assistants and agents remember over time?"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2410.10813] LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory, CC BY 4.0; arXiv, [2602.22769] AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was rerun; only the fetched arXiv abstract pages were read. Results are as reported by the authors."
category: agent-operations
topic_area: "agent operations: memory"
tags: agents, memory, benchmarks, long-term
sources:
  - id: S1
    title: "[2410.10813] LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory"
    publisher: arXiv
    url: https://arxiv.org/abs/2410.10813
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2602.22769] AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications"
    publisher: arXiv
    url: https://arxiv.org/abs/2602.22769
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# How well do assistants and agents remember over time?

## Short answer
Not well yet. LongMemEval saw a 30% accuracy drop for remembering across sessions. AMA-Bench found memory systems fail on agent trajectories because similarity retrieval loses causal detail. [S1][S2]

## Design hints
- Split memory into indexing, retrieval and reading; index by facts and time. [S1]
- For agents, store causal links between steps, not just similar text. [S2]

## Claims
- id: C1
  claim: Assistants show a 30% accuracy drop across sustained interactions.
  source: S1
  evidence: "showing a 30% accuracy drop on memorizing information across sustained interactions"
- id: C2
  claim: Memory design has three stages.
  source: S1
  evidence: "three stages: indexing, retrieval, and reading"
- id: C3
  claim: Existing memory systems rely on lossy similarity retrieval.
  source: S2
  evidence: "rely heavily on lossy similarity-based retrieval"
- id: C4
  claim: Existing benchmarks are dialogue-centric.
  source: S2
  evidence: "existing memory benchmarks are largely dialogue-centric"

## Sources
- [S1] [2410.10813] LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. arXiv. https://arxiv.org/abs/2410.10813 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2602.22769] AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications. arXiv. https://arxiv.org/abs/2602.22769 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).