What a Persistent Trading Agent Must Remember Between Conversations
By DX Research Group · · DXAP platform
A worked lifecycle for persistent agentic trading: separate mandate, market state, and settled history before the next decision.
A useful trading conversation can finish while the trade remains open. That creates an engineering question: what carries the strategy forward after the user closes the browser? DXAP describes a persistent agent. We think persistence earns its value through continuity of authority and state, rather than through the length of a conversation.
Three records with different jobs
Consider an illustrative strategy: observe an asset, enter after a confirmed recovery, and leave when the original thesis breaks. A chat answer can describe that plan. A persistent runtime must continue interpreting it as prices and balances change. We would inspect three records separately.
The mandate contains the owner's current instructions and constraints. Market and portfolio state describe the account now. Settled history records what earlier actions actually changed. Mixing these records creates predictable mistakes: an old proposed entry looks like a current holding, or a past explanation looks like a permanent rule.
Our controls paper companion provides the architectural background for keeping authority and recorded decisions distinct. The implementation question here is narrower: can the next turn reconstruct those three records without relying on the user's memory?
Follow one position across a restart
At the first observation, the recovery condition remains incomplete. The agent records an observation and continues waiting. Later, an entry becomes eligible. The runtime checks current account state before submitting the proposed order. Following a fill, the held quantity comes from the venue record, while the thesis continues to come from the mandate.
Suppose the runtime restarts before the next scheduled review. A useful persistence test asks it to rebuild the same current position and active instructions. It should distinguish the filled quantity from any unfilled remainder. It should also recognize whether the user revised the strategy while it was offline. Replaying an old conversation verbatim would miss both transitions.
This scenario is a design test, rather than a report of a particular production incident. It shows why an agent's continuity depends on recoverable records. Repeated inference alone cannot supply continuity when the input state is wrong.
Evaluate continuity before market judgment
We would test persistence with a frozen market sequence and an interruption inserted between order acknowledgement and the next turn. Compare an uninterrupted run with the interrupted run on restored positions, active mandate version, and unresolved orders. Only after those agree does it make sense to compare the next decision.
A second fixture changes the mandate during the interruption. The resumed agent should use the new instruction while preserving the historical reason for the earlier action. This separates strategy evolution from accidental rewriting of history.
The continuous research record explains why behavior and economic outcomes need their own measurements. Persistence is an architecture advantage when it reduces the work needed to maintain an ongoing strategy and makes transitions inspectable. A reader evaluating agentic trading should ask to see a restart, a changed mandate, and a reconciled account. Those concrete demonstrations reveal more than a long memory transcript.