LAB-0019 / ACTIVE
Long-Term Agent State
Testing which events deserve durable storage and which should decay after task completion.
Experiments are where the archive is allowed to be wrong in public. Each lab has a question, a setup, observable failure modes and a link back into the concept graph.
SORT: STATUS / UPDATED
Testing which events deserve durable storage and which should decay after task completion.
Combining lexical, vector, metadata and graph retrieval under one routing policy.
Routing extraction, classification and formatting tasks away from large models when confidence is sufficient.
Studying idempotency, retry budgets and error classification for model-driven tool calls.
Evaluating summaries, structured state and selective retention under constrained windows.
Using graph neighborhoods as additional retrieval candidates for concept-heavy questions.