03 / MEMORY & RETRIEVAL
Find what matters now.
Keep the original record.
Vectors can help locate relevant information. After a match, retrieve the original text, source and constraints to judge whether a memory applies.
Interactive illustration: all records are fictional, coordinates are hand-set in two dimensions, and ranking uses Euclidean distance. No embedding model is called and your memories are not stored here.
Change the question; inspect the matches
Diamond: query · bright dots: matches · dashed: old version · gray M3: private
Give the agent source records
Nearby does not mean correct. More matches increase context length and may also introduce irrelevant or conflicting material.
Three distinct questions
| Question | Direction to explore | Still needed |
|---|---|---|
| Find older information? | Vector indexing and similarity search | Sources, versions, time and access control |
| Reduce the input? | Retrieve relevant passages, then test phrase compression | Missed information, comprehension and full costs |
| Can agents exchange vectors directly? | Research shared encodings and receiver adapters | Compatible spaces, models and interfaces; arbitrary models do not automatically interoperate |
Vectors should not be the only copy
Vectors commonly serve as retrieval indexes without promising exact reconstruction. Retain source content and permissions separately. Embeddings are not anonymization: research has demonstrated text-recovery attacks under specific conditions.
Questions for the next experiment
How should old memories expire after a change? How do two agents confirm the same meaning? Can retrieving fewer passages cause more retries? Each needs its own measurement.
Join the memory discussion ↗