Search papers, labs, and topics across Lattice.
This paper investigates the concept of memory depth in long-running language agents, emphasizing the importance of durable goal-conditioned tendencies over mere memory access. The authors introduce the loop-drift protocol to evaluate how well agents maintain behavior after unloading working context, finding that their proposed EVAF mechanism significantly enhances goal persistence and recovery compared to traditional retrieval methods. Results show that while retrieval excels at factual recall, EVAF achieves superior performance in maintaining goal-oriented behavior with minimal parametric writes, highlighting the necessity of selective consolidation in memory systems.
EVAF achieves up to 90% goal persistence in language agents, revealing that memory depth is crucial for sustained behavior beyond simple retrieval.
Long-running language agents need more than memory access. Retrieval systems can fetch past facts at query time, but they do not decide which experiences should continue to shape behavior after the working context is unloaded. We study this separate problem as memory depth: durable goal-conditioned tendencies written into a small parametric store. We introduce the loop-drift protocol, a controlled stress test in which the retrieval index remains intact while working context is unloaded and goal-conditioned behavior must persist under long-loop interference. We evaluate EVAF, a surprise- and valence-gated LoRA consolidation mechanism. Across GPT-2 and TinyLlama, retrieval is strongest on shallow factual recall (short-fact accuracy 0.956--0.973), while EVAF is strongest on goal persistence and post-unload recovery (0.812--0.904) with only 2--3 parametric writes per 200 events. Mechanism controls show that selective consolidation factorizes into two controllable dimensions: selection and actuation. Matched random gates isolate selection beyond sparse writing; fixed-inner controls across GPT-2, TinyLlama, and Mistral-7B show that inner-loop write strength is model-dependent; and a Mistral-7B matched-gate inversion reveals asymmetric selection-actuation coupling under miscalibrated actuation. Public Memora event streams serve as an external diagnostic, exposing stale-memory invalidation as an unresolved boundary. Within this probe, selective parametric consolidation supplies memory depth distinct from and complementary to retrieval access.
