01 / Strategy records
Keep the thesis next to the trade
Save the idea, its entry conditions and what would prove it wrong as separate records. The agent recalls them before it proposes anything.
ThesisEntry ruleInvalidation
Strategy that survives a restart
Memory for trading agents.
A trading agent that restarts every morning forgets why it liked a setup yesterday. Pinlo keeps the thesis, the risk limits you set and the reasons behind each past call, so the next review starts from them.
Example
Desk agentRequest
A breakout alert just fired on the pair we watch. Check it against what we agreed last week.
Recalled from the mind3 records
- 01ThesisOnly trade the breakout if volume holds above the 20-day average for two candles.
- 02LimitsMax 2% of the book per idea; anything larger waits for a human approval.
- 03Last callPassed on a similar alert on Tuesday: volume faded within the hour.
- 01Read the alert
- 02Recall thesis and limits
- 03Write the review
Illustrative request and records, made up to show where memory fits.
The problem
What slips away
between sessions.
01
The reasoning resets
After a restart the agent sees the price and the alert, but not the conditions that made the idea worth watching in the first place.
02
Limits get typed in again
Position sizes, approval rules and banned instruments have to be pasted into every new session, and one missed paste is one unchecked trade.
03
Logs without the why
An exchange fill says what happened. It does not say which research, which rule or which earlier rejection shaped the decision.
With Pinlo
What this agent's memory holds.
02 / Tag recall
Pull only what the signal needs
Tag records by pair, strategy and event. Recall filters by tag and keyword, and your code decides which matches go into the prompt.
pair:eth-usdgstrategyreview
03 / Decision trail
Write down every call
Store the observation, the reasoning and the outcome as three records. Each one gets a hash, and the version you anchor on Robinhood Chain fixes the set.
ObservationReasoningOutcome
04 / Shared desk
Let a second agent check the first
Grant a review agent read access to the strategy mind. It reads the same limits the trading agent wrote, without a copy to keep in sync.
Research agentReview agentRead grant
The workflow
How the work
uses memory.
- 01
Record the signal
The agent writes the alert, its source and a timestamp, tagged with the pair and the strategy it belongs to.
- 02
Recall before acting
Before proposing a trade it recalls the thesis, the limits and the last few calls on that pair, and puts them in the request.
- 03
Save what happened
The chosen action, or the reason for passing, is saved with the outcome once it is known. The next anchor version includes it.
Where the line is
Pinlo stores and proves records. It does not place orders, enforce limits or judge a trade: your trading code still checks every limit before an order leaves it. A matching hash proves a record was not changed, not that the trade was a good one.
- Open one mind per strategy or per account, so recall never mixes two books.
- Tag every record with the instrument, the strategy and the kind of event.
- Orders, permissions and limit checks stay in your trading service, outside the model.
- Store source links and timestamps with every summary the agent writes.
Questions
Worth asking
first.
Does Pinlo place trades?
No. Pinlo holds the agent's memory. Orders, wallets and risk checks stay in your own trading service.
Will memory make an agent trade better?
Nobody can promise that. Memory means the agent sees the rules and history you saved; whether its calls are good still depends on the strategy and on your checks.
What should a trading agent save?
The thesis and its invalidation, sizing and approval rules, each alert it reviewed, the reasoning it gave and the outcome when it is known.
What can someone verify later?
That a record was part of an anchored version: its hash and Merkle proof against the root stored by PinloCortex on Robinhood Chain. That proves integrity, not accuracy.
Related use cases
Forecasting agentsEvidence, odds and revisions kept in order for each question.Asset researchDocuments, versions and issuer updates for tokenized assets.Field robotsA proposed memory layer for devices, sites and fleets.