Memory Maintenance & Pruning

Memory stores grow. Without maintenance, they become bloated with outdated episodic memories, redundant semantic entries, and orphaned references. Memory maintenance is the operational discipline of keeping the store lean, accurate, and fast.

The maintenance cycle has three operations. Pruning: remove memories that have not been accessed in 180 days and have no forward references. These are memories that the agent never retrieves — dead weight that slows search and consumes storage. Consolidation: merge redundant semantic memories. If the agent has learned "Client X prefers concise reports" three times from three different interactions, consolidate into a single memory with higher confidence and three source references. Verification: for memories older than 90 days that are still actively retrieved, validate against current reality. Is the client contact still in the same role? Is the pricing data still current? Is the market assumption still valid?

1. 1. Prune Monthly Run a monthly pruning cycle: identify memories with zero accesses in the last 180 days and no forward references. Archive them to cold storage — do not delete. The archive is available for manual retrieval if needed.
2. 2. Consolidate Quarterly Identify semantic memories that express the same knowledge with different wording. Merge into a single entry with the highest confidence score and all source references preserved.
3. 3. Verify High-Access Memories For the top 20% most-accessed memories, run a quarterly verification: is this still true? Flag unverified memories. Update verified memories with a fresh timestamp. This prevents the most-used memories from being the most stale.