AGENTCL: Toward Rigorous Evaluation of Continual Learning in Language Agents

Why Your Agent's Memory Might Be Hallucinating Its Own Past

Current AI agents spend significant time solving complex tasks. However, they rarely get smarter from the experience. They treat every new problem as a fresh start. They fail to accumulate reusable knowledge from one episode to the next. Current benchmarks struggle to test if an agent is actually learning. They mostly check if an agent can find a needle in a haystack of old conversations. This paper introduces AGENTCL. It is a framework that creates "compositional" task sequences. In these sequences, solving an easy task first provides the building blocks for a harder one later. This allows us to see if an agent's memory captures transferable skills or just creates noise.

The Question

Can we rigorously measure whether a language agent is performing continual learning? Continual learning is the ability to accumulate reusable experience across a stream of tasks. Can an agent do this without benefiting from accidental task overlap? The authors investigate non-parametric memory designs (systems that store and retrieve experiences without updating the model's underlying weights). They look at the tension between plasticity (the ability to adapt to new tasks) and stability (the ability to retain past knowledge without corruption).

Why The Old Answer Was Incomplete

Until now, the field has relied on two flawed proxies for learning. The first category involves long-context benchmarks. These test if an agent can retrieve information from a massive document. These are static. They test retrieval, not adaptation. The second category involves "lifelong" benchmarks that stream tasks sequentially. However, they lack control over the relationship between those tasks.

As shown in, most existing benchmarks use "naive" streams.

These pull tasks from a domain at random. In these settings, if an agent performs better on Task 2 than Task 1, we cannot prove learning occurred. We cannot tell if the agent learned a reusable workflow. Perhaps Task 2 was just easier or similar in domain. This lack of structure compresses the performance differences between memory architectures. It makes it nearly impossible to distinguish a sophisticated memory system from a simple one.

What They Did

The researchers developed AGENTCL to force a distinction between coincidence and competence. They constructed "compositional" task streams. In these, earlier sub-solutions or pieces of evidence are explicitly required for later, more complex tasks . They evaluated several non-parametric memory designs. These ranged from simple retrieval systems like Mem0 to structured approaches like Agent Workflow Memory (AWM).

To diagnose these systems, the authors implemented MEMPROBE. This is a probing method using a retrieve-solve-consolidate loop. MEMPROBE decomposes memory into three distinct layers. It uses interaction memory (raw trajectories), insight memory (abstracted patterns and failures), and skill memory (procedural snippets). They ran these through a two-pass evaluation protocol .

The first pass allows the agent to both read from and write to memory. The second pass uses a frozen memory to see if the "learned" knowledge actually persists.

What They Found

The results suggest current benchmarks may hide the true limitations of agent learning. The authors report that compositional streams provide much higher discriminative power than naive ones. For example, on the CodeEval-Pro dataset, the standard deviation of accuracy across methods was 9.4 in compositional streams. In naive streams, it was only 3.0 .

This means compositional tasks make it much easier to see which memory designs actually work.

Crucially, the study reveals a massive stability bottleneck. Some methods show impressive Plasticity Gain (PG)—the improvement over a memoryless baseline. ExpRAG and MEMPROBE achieved +17.7 and +21.9 percentage points respectively on CodeEval-Pro . However, their Stability Gain (SG) is often zero or negative. Negative SG means the "learned" knowledge was lost or corrupted by the time the next task arrived. Even more concerning, memory can induce degradation in held-out settings. In these cases, the accumulated memory acts as interference rather than assistance.

What This Changes

If these findings hold, the industry's focus on "bigger context windows" may face diminishing returns.

First, for practitioners building long-running agents, memory cannot be a passive dump of past logs. Simply storing more trajectories will eventually lead to "cognitive interference." This happens when an agent is confused by semantically similar but functionally incompatible past experiences. Effective memory requires active distillation. You must decide what to abstract into a "skill" and what to discard as noise.

Second, for researchers, this shifts the goalpost. The goal is no longer just maximizing "retrieval accuracy." Instead, you must manage the plasticity-stability tradeoff. We need architectures that do not just remember, but selectively forget. The paper demonstrates that a memory system highly plastic in a controlled environment can become a liability in the wild.

Figures from the paper