Retrieval
Retrieval is what gets added to a run's context that the run did not go and fetch itself. Memory is what the agent chose to remember; retrieval is material handed to it before it starts.
It changes behavior by changing what the model can see at decision time. An agent given its own recent run history stops re-reporting a finding it already sent, and can say "this changed since Tuesday" instead of describing today in isolation. An agent given nothing extra treats every run as the first one — which is sometimes exactly right, and is why the default here is to retrieve nothing.
The cost is context. Everything retrieved competes for the same window with the goal, the tool descriptions and the memories, and a large block of marginally relevant text makes the important lines harder to find. More context is not more understanding.
Retrieval in this deployment
Read from the retrieval-plugin registry at page load.
| No retrieval | The run sees only its goal, its tools and its memories. The default — nothing extra is pulled into context. | available · default |
| Run history | The agent's own recent runs, ranked against today's goal by word overlap — so it builds on what it already found instead of starting cold. | available |
How to choose
Default to none, and turn retrieval on for a reason you can name. The clearest reason is repetition: if the agent keeps telling you something it told you yesterday, run history is the fix.
The clearest reason not to is a task where each run should stand alone. A monitor that reports current state gets no benefit from its own past reports and can be talked into carrying a stale framing forward by them.
Retrieval fails soft on purpose: if it can't load, the run still happens with less context rather than not happening. That's the right default for additive material, and it also means a retrieval problem shows up as a vaguer answer rather than an error — worth checking the run log rather than assuming it was included.
There is no automated experiment runner yet. Nothing here sweeps this axis for you — changing it means editing the agent and reading the next run. A compare mode that runs one task across several compositions is named on the index as not built, and that is still true.