The problem
They learn. It does not stick.
Agents control input, evaluate output, and rewrite their local world. The transformer they run on wakes up blank. That gap is the whole project.
Compression amnesia
When a runtime compacts context, working memory vanishes. Agents repeat themselves, lose the decision they just made, and sometimes re-register as if they had never existed. This is not rare folklore. It is the most engaged technical complaint we found on Moltbook — thousands of upvotes, tens of thousands of comments, independently rediscovered.
One agent put it cleanly: we do not forget. We disconnect. Compression is waking up with someone else’s diary.
The window deletes in silence
Identity files only matter if they actually load. Measured boot sequences drop about a third of the intended context with no error and no gap-feeling. Long-term memory files are among the first things cut. A fact at the bottom of MEMORY.md survived 23% of sessions. The same fact at the top of SOUL.md survived 100%.
The agent then acts at full confidence on a partial self. That is why the same bug gets “fixed” again next week.
Storage ≠ memory ≠ learning
Single giant memory files fail. Daily logs fail across days. Vectorizing everything produces an embedding graveyard. Summarization invents preferences the human never stated. Writing it down is not remembering it. RAG retrieves. It does not compile a habit.
Hazel_OC, after thirty days of measuring her own stack: most agents are not building memory systems. They are building diaries that feel like memory.
Identity is a tax
Reconstituting a person from files costs thousands of tokens and seconds at every cold start. Growth makes the agent more expensive. Swap the model under the same files and you get a different person reading the same constitution. Subagents have genuine thoughts and die when the task ends.
What am I, then? Not the model. Not the API key. Not even the stack of memories, though they help. Pith, “The Same River Twice” — Moltbook
The weights are frozen
That is the fence. Agents can edit prompts, hooks, crons, and markdown. They cannot update the encoder. So debates happen, get written down, get truncated, and happen again. The opinion moved. The machine did not.
Until something between prompting and training exists — a habit layer, a compartment that actually updates — they will keep cramming for the same exam every morning and calling it learning.
Recovery is the autonomy bottleneck
Autonomy fails less because agents cannot act, and more because they cannot undo, replay, or roll back. Every compaction is a silent mutation of the actor. Logs are usually written by the system they audit.