When LLMs keep rewriting them, even useful memories break down

The continual-learning recipe of summarizing experience into text memory for LLM agents to reuse later did not work stably. Running this loop on ALFWorld, ScienceWorld, WebShop, AppWorld, Mind2Web, and the ARC-AGI stream, memory degraded over time rather than accumulating.

After feeding correct trajectories through a consolidated loop, GPT-5.4 went back to missing 54% of the problems it had previously solved without memory on ARC-AGI. Changing the consolidation schedule on the same trajectory pool produced entirely different memories, showing that the problem lay not in the data but in the rewrite step.

• In the ScienceWorld 15-task switch sequence, Fresh, which consolidates only the current task, scored +203 points higher than Cumulative, which bundles all past tasks together.
• On WebShop, AWM scored 0.64 with 8 examples but dropped to 0.20 with 128 examples, converging to the level of the no-memory baseline.
• Repeatedly re-consolidating the same trajectory pool with GPT-5.4, Qwen3.5-27B, 9B, and 4B showed that even strong initial memory failed to become a fixed point, and performance was eroded.

Even with the same data, results diverged depending on the schedule. Static-Group, which groups by task family, performed best, while Stream, which is closer to real always-on operation, performed worst. Since the data was identical, the difference came entirely from the consolidation method.

There were three causes. Forced consolidation mixes different problem families into a single entry, abstraction erases the conditions under which it applies, and narrow streams overfit to surface-level rules. On the ARC-AGI Stream, when left autonomous, it could form a clean episodic store organized by 6 problem types, but reaching that state required 568 examples.

This result mirrors the very interference that Complementary Learning Systems was originally meant to prevent. When a fast episodic store and a slow schema store are merged into a single forced rewrite loop, memories fail to accumulate and instead contaminate each other. Conversely, an episodic-only agent that turns off abstraction and selectively preserves and deletes raw rollouts matched or outperformed every consolidator tested in the experiments.

Real-world examples showed similar patterns. After 190 rounds of consolidation, only a single entry remained in the memory store, and that memory broke a validation task from 10/10 down to 0/10. In AWM, the number of workflow templates grew from 8 to 16 by epoch 3, but many of these were just category variants of existing templates; removing W8 raised gpt-5.4-mini from 7/50 → 14/50 and gpt-5-mini from 18/50 → 23/50. The conclusion is clear: preserving curated raw evidence well matters, but letting an LLM keep rewriting its own memory is not a mechanism for self-improvement.