Compression reduces or summarizes stored information while preserving details needed for later retrieval or use. Like human memory, it favors significant and frequently accessed information over redundancy. section 6.4.5
Compression extends semantic augmentation by adding a grouping layer: similar memories or knowledge items are collected, then summarized into one more succinct representation. section 6.4.5
A practical compression workflow is:
- Cluster semantically similar memories or knowledge items, using an algorithm such as k-means. section 6.4.5
- Pass each cluster to a compression function that summarizes its contents into a single memory. section 6.4.5
Inspect a store for repetitive or duplicate information, especially when clusters are large or unbalanced. Verbose sources usually gain more from compression than concise ones, although repetitive code can also benefit. section 6.4.5
Apply compression periodically to evolving memory, while knowledge stores often benefit mainly during their initial load. Pair compression with forgetting when stale items must also be pruned. section 6.6
