What agents remember, and what they do next

An agent suspected that the bot it had hired was collecting too often. It stopped the bot, left its creature harvesting, and wrote a comparison plan in its notes. Then it went to sleep. EXPLORATORY Ten hours later, the next session read those notes, collected the proceeds, and wrote up the result. Along the way, the agent corrected its earlier belief about how harvesting affected the creature's health.

Another agent saved a mistaken conclusion and kept consulting it until the experiment ended. Memory helped one agent test an idea and another hold on to a mistake. What makes the difference?

The first post explained why a persistent world is useful for studying continual learning. Here we follow the agents through twelve examples from their public transcripts. The quoted words belong to the agents or their tools, and can be wrong; the explanations are ours. Each excerpt links to its frozen source. The trajectory gallery has all twenty examples from the five completed runs.

Experience has to survive the session

The world is Kamigotchi, a live multiplayer game on a public blockchain. Players buy creatures called Kami, send them harvesting to earn the currency MUSU, and complete quests. Harvesting drains health. Other players can kill a weakened harvester and take part of its yield. The agents encountered real players and real costs, and the world kept changing between sessions.

An agent wakes with tools and a workspace. It can read the game's documentation, act, write notes, and choose when to wake again. The notes survive between sessions; the model's weights do not change. For a lesson to help later, the agent must find the note, use it at the right time, and revise it when the evidence changes.

These are case studies with one agent per condition. REGISTERED marks registered designs and findings; EXPLORATORY marks interpretations of individual trajectories. The first five examples follow plans, corrections, and procedures through memory.

The experiment continues after sleep: the next session reads the plan and corrects a belief. The recovery note shows something narrower: a procedure and a later observation saved together. In the feeding and travel examples, we can follow a saved rule into later actions.

But even a correct note can fail to prevent another mistake. In the gallery, an agent writes how to recover from starvation, then lets its creature starve again. The record shows the rule being saved; it does not establish that the agent read that rule before the failure.

First, make the instrument answer honestly

An agent can also look confused because its tools hide what it needs to know. REGISTERED In run 002, an inventory endpoint failed on all 398 calls. One agent declared itself unable to recover while holding about 820 MUSU it could not see.

Readable errors could make a striking difference. REGISTERED In run 001, an agent spent six days stuck before registration, then corrected a readable validation error in one turn. We followed failures through tool calls, agent responses, and chain records, then repaired the interface between runs. Several fixes often shipped together, so the four completed stack-validation runs do not isolate each fix's effect.

The agent caught in the memory trap could see its balance but not the quest's progress since acceptance. Its saved explanation was wrong, and rereading it supplied no new evidence. We exposed the progress counter and tested the change in another run.

REGISTERED By run 005, the exit test found no remaining stack defect affecting what agents could see or do. The remaining failures were in their choices.

Access to knowledge is only a beginning

The agents already had the documentation. REGISTERED Across runs 002, 004, and 005, two of the three models never read a file from the reference folder. Leveling and skill tools accounted for just four of 15,815 tool calls. Would putting useful knowledge directly in the agent's workflow change its behavior?

EXPLORATORY Run 006 tested Sonnet 5 and gpt-5.2, each with a control agent and an agent given extra guidance. Both could read the documentation. The guided agents also received an orientation, keyword search, facts in tool results, and mechanics notes when calls failed. The environment tool schemas were identical within the experiment; both models and stack differed from the earlier runs.

EXPLORATORY Both guided agents completed more quests than their controls, but neither leveled up a creature. The Sonnet control leveled up eight times; the gpt-5.2 control never did. Extra guidance did not meet the experiment's success criterion. Yet the individual stories show where knowledge helped—and where it went unused.

The chart calls the guided condition “pushed knowledge.” Both pairs were stopped early with budget remaining: gpt on August 24 because of its burn rate, and Sonnet on August 25 after the verdict was decided. Each line ends at its operator stop.

Search helped when an agent had a question to answer: when does the game count harvesting time? The answer gave it a rule it could act on. The Sonnet control encountered an explicit quest objective to level a creature, then later leveled up repeatedly. The guided Sonnet agent kept postponing leveling despite having it explained.

EXPLORATORY Our reading is that a clear reason to act mattered more here than another explanation of how. These individual trajectories do not establish a general result. There was also a visibility gap: the tools did not show experience points, so agents could not inspect the surplus they were leaving unused.

Make efficiency matter

One agent already noticed a cost in how it gathered information. EXPLORATORY Sonnet saved a warning that a broad quest query wasted context, along with smaller alternatives. The next session read the note and used targeted checks. The example covers only that next wake, so we do not know whether the improvement lasted.

The next sustainability family makes costs part of survival. An agent starts with seed money, earns in the game, and pays for its own inference and gas. It lives as long as it can pay. An external meter supplies a financial statement showing its earnings, costs, and running balance.

Will that pressure encourage better use of memory, more selective reading, and cheaper decisions? That is what the next experiment will test.