I tested 3 different memory architectures for long running agents,here’s what actually broke

After,building agents,that need to maintain context. across multi-day tasks, I found,that memory design is where most projects quietly fail. What I tried: Full conversation history in context - Works until it doesn’t. Token limits hit faster than expected, and older context gets “forgotten” even when technically present. Retrieval quality degrades. Vector store retrieval only - Cleaner, but agents kept losing track of their own decisions. They’d contradict earlier reasoning because episodic memory wasn’t structured separately from factual memory. Hybrid: structured episodic log + vector retrieval - This held up best. The agent writes its own decision summaries to a structured store, then retrieves both facts and past reasoning separately. One thing that made the hybrid model click for us: when agents need to refresh beliefs mid-task by pulling live external data, unreliable fetching poisons the memory store fast. We ended up routing those external lookups through residential rotating proxies so the agent gets clean, unblocked data on every retrieval cycle - bad input is a silent memory killer that’s easy to miss. The real lesson: agents need memory that mirrors how they think, not how humans archive information. Biggest open question I’m still wrestling with: how do you handle memory conflicts when an agent updates its beliefs mid-task? Do you overwrite, version, or flag for human review? Curious what architectures others have landed on, especially for agents running longer than a single session.