Simsek et al. introduce Memory Worth (MW), a two-counter per-memory signal tracking co-occurrence with successful versus failed outcomes, converging almost surely to p+(m) = Pr[y_t = +1 | m in M_t] under stationary retrieval and minimum exploration (arXiv:2604.12007, 2026). Primary source establishes MW as associational rather than causal, enabling staleness detection, retrieval suppression, and deprecation where prior static write-time scores or LLM heuristics fell short. Coverage in related works such as Reflexion (Shinn et al., arXiv:2303.11366) emphasized verbal reinforcement loops but omitted lightweight outcome-feedback primitives for long-term memory quality.

Generative Agents (Park et al., arXiv:2304.03442) demonstrated believable simulacra via memory streams yet suffered from unchecked accumulation, a scalability failure MW directly targets for persistent agents operating across shifting task distributions over thousands of episodes. Original paper validates rho = 0.89 +/- 0.02 Spearman correlation after 10,000 episodes in synthetic environments and threshold behavior (MW = 0.17 for stale, 0.77 for specialist) with all-MiniLM-L6-v2 embeddings; prior coverage missed explicit linkage to lifelong autonomy where memory bloat becomes prohibitive.

This governance primitive addresses the core scalability challenge for long-term autonomous systems by providing a theoretically grounded, constant-space mechanism that prior retrieval-focused architectures neglected, synthesizing outcome logging already present in most agent frameworks with proven convergence guarantees.