Memory Decay
Authoritative Definition
Algorithmic recency weighting and garbage collection strategies applied to historical memory records to prioritize current context and manage state size.
Overview & Technical Description
Memory decay mechanisms dynamically reduce the retrieval ranking score, or entirely prune, stale historical items based on elapsed time or relevance decay formulas. It operates on the principle that more recent interactions typically hold higher contextual importance than older, potentially outdated information. In complex agent systems, memories are often assigned a decay function (e.g., exponential decay) that lowers their activation probability over time unless they are repeatedly reinforced or recalled. This helps the retrieval system naturally surface the most current user preferences and environmental states, mitigating the risk of the agent acting on obsolete facts or conflicting historical intents. Beyond scoring adjustments, physical memory decay acts as a garbage collection process, archiving or deleting low-value context tokens to maintain optimal database performance and reduce vector search latency. This lifecycle management is critical for agents operating continuously in production environments over extended periods.
Editorial Notes
Ensures recent user context overrides outdated state. Fine-tuning the decay rate is highly application-specific; overly aggressive decay can cause an agent to exhibit amnesia for important but rarely accessed facts.