Agent memory comparisons and data
Investigate how retained information reaches later work, changes over time and affects behavior. The Agent Memory topic introduces the questions; this directory holds comparative research and dataset snapshots, not copies of the individual analyses.
Comparative research
- How memory reaches the next decision — selection, delivery and consumer identity across contrasting routes.
- How corrections propagate through retained memory — invalidation, withdrawal, index freshness, derived prose and stale consumers, with experiments for testing propagation.
Agent-memory landscape compares admission,
read-back, curation and trace learning across 243 analyses. This is a frozen
synthesis dated 2026-11-09 at inputs commit bf2da04db, not a continuously
updated survey or a representative sample of the field.
Datasets and individual systems
- Memory comparison table
- CSV download
- Current memory accounts and profiles — generated during each site build, linking the accepted memory members without copying them.
The table and CSV were generated on 2026-11-09 from all 243 analyses current at
commit bf2da04db: 225 code-grounded and 18 doc-grounded sets. Their population
can differ from the build-time listing.
A row records a scoped finding, not a blanket property of the whole system.
Read evidence and coverage alongside the value.
Method
Methods & Data explains evidence interpretation. Comparison inputs and regeneration owns population selection, profile revision semantics and rebuild commands. Cross-system syntheses follow the separately commissioned landscape synthesis procedure.
The memory analyst job traces retained objects and their later consumers in each frozen analysis.
All comparisons · Research Topics
Complete file listing (generated at build time)
Current analyses (generated at build time)