Concentrated liquidity market makers (CLMMs), such as Meteora's Dynamic Liquidity Market Maker (DLMM) on Solana, require liquidity providers to continuously select pools, size discrete bin ranges, and time entries and exits. This constitutes a sequential decision problem under partial observability that static, parametric rebalancing vaults (e.g., Kamino Finance, 2024; Gamma Strategies) handle poorly in volatile, high-turnover markets. Conversely, recent large language model (LLM) trading agents (e.g., FinMem (Yu et al., 2023), FinAgent (Zhang et al., 2024)) focus predominantly on isolated single-agent spot or perpetual trading, lacking domain-specific bin mechanics and cross-instance knowledge sharing. We present Etemaro, an autonomous architecture that utilizes an LLM inside a ReAct reasoning loop to screen candidate pools, deploy and manage DLMM positions, and execute exits based on live profit-and-loss, yield, and bin drift data. Etemaro introduces three primary contributions beyond standard trading-agent frameworks: (1) a Darwinian signal-weighting mechanism that reweights heuristic screening signals online using a quartile-lift statistic computed from realized win/loss outcomes; (2) role-scoped and intent-filtered tool access, strictly constraining which on-chain actions the LLM may execute across screening, management, and chat contexts; and (3) HiveMind, a lightweight federated experiential-learning protocol where independent agent instances exchange natural-language lessons and scored strategy presets without sharing raw transaction logs, private keys, or model weights. We evaluate Etemaro through a continuous 114-day production deployment on Solana Mainnet (April 1 to August 24, 2026). Across this period, the agent executed 4,583 on-chain deployments committing 9,715.52 SOL in capital (~2.12 SOL average per position) and generated 50,329 on-chain lifecycle events, achieving a 68.35% realized win rate (743 wins out of 1,087 closed positions) with a median hold time of 36.5 minutes. We candidly document operational failure modes, including a 37.25% state divergence between local logs and on-chain reality caused by RPC dropouts and 5,521 concurrent race-condition transactions. We position this work as a systems and architecture contribution, establishing an empirical template for safety-constrained, self-improving autonomous agents operating in live financial environments. Overall, our findings demonstrate that combining bounded reasoning loops with lightweight experiential federation enables autonomous agents to operate effectively and self-adaptively in high-velocity decentralized financial markets.