Normalized Telegram record
Agent memory is being evaluated as system behavior
MemoryData evaluates what an agent stores, retrieves, updates, and forgets across a sequence instead of grading one final answer.
Signal contract
MemoryData evaluates what an agent stores, retrieves, updates, and forgets across a sequence instead of grading one final answer. Useful memory is a lifecycle with write and invalidation decisions, so point-in-time retrieval metrics cannot establish whether an agent learns safely over time. Novelty: structural. Verification: source-inspected. This New Runtime record is an evidence-linked retrieval unit.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | huggingface.cosource | primary receipt | source_urls |
| 2 | github.comrepo | supporting receipt | source_urls |
Observation
MemoryData evaluates what an agent stores, retrieves, updates, and forgets across a sequence instead of grading one final answer.
Why it matters
Useful memory is a lifecycle with write and invalidation decisions, so point-in-time retrieval metrics cannot establish whether an agent learns safely over time.
Entities
MemoryData
Provenance
This public record is an English normalization of QWG AI Telegram message 2625. The complete original-language post remains the canonical raw message.