LetTokenBurn
Sign in
Auto57%Harbor: pendingAgentic / Tool Use

PraxisMemBench

PraxisMemBench: Benchmarking Personal and Procedural Memory for Persistent Assistants

A harness-based benchmark evaluating persistent personal assistants on coupled personal and procedural memory across multi-session life and work trajectories.

Paper

How to run

ltb · praxismembench
$ npm i -g lettokenburn-cli
$ ltb bench info praxismembench

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

4 sourced entries
HarnessSourceConfidenceLink
DeepSeek V4 Pro49.8N/AIndependentmediumOct 14, 2026arxiv.org
GLM-5.246.7N/AIndependentmediumOct 14, 2026arxiv.org
Qwen3.7 Plus43.1N/AIndependentmediumOct 14, 2026arxiv.org
PraxisMem37.9N/AIndependentmediumOct 14, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

PraxisMemBench constructs sequential persona trajectories spanning daily-life and work domains, where 84.2% of tasks require both personal context and procedural skills. It evaluates how agents maintain and retrieve evolving user preferences and application workflows over time.

Scoring uses Session Success Rate to measure task completion. Experiments reveal that maintaining memory types separately is insufficient, with performance dropping on interdependent tasks, highlighting a continual-maintenance challenge in persistent assistance.