- Who it helps
- Developers evaluating memory for personal assistants.
- Availability
- The October 25 paper is public. Its linked project page was unavailable during this check; released artifacts were not verified.
Measure losses as well as gains
The authors compare dedicated memory systems with retrieval from raw history. They examine faithfulness, accessibility and robustness: a memory can help one query while obscuring evidence needed by another.
A deadline check
An original evaluation example: change a deadline from Friday to Monday, then ask for the current date and the passage supporting it. Compare the stored-memory answer with retrieval from the original messages.
What you can do
- Compare answers and their supporting passages on the same histories before choosing a memory design. Open reference ↗
Where the evidence stops
These findings do not establish that every memory system performs worse. The benchmark results were not reproduced.
Sources read. Public sources checked October 28, 2026. No installation, model inference or complete product workflow was executed.
Sources & further reading
Paper · paper
Memory Tax: when stored memories lose useful evidence ↗
Supports the subject identity, stated behavior and verification limits.
arxiv.org · page checked 2026-10-28
Coverage note: Original X posts were inaccessible; comparable independent demonstrations, discussion and complete historical coverage were not established.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Memory Tax examines whether an agent’s memory system loses evidence that direct retrieval from the original history could preserve.