Strategy thesis
Record the conditions that would support or invalidate the idea.
TRADING AGENTSEvery trade.
Every trade.
The full context.
Carry strategy, risk rules and decision history into the next session. Give your agent the context behind its next move.
Observe→Recall→Record
MEMORY IN PRACTICE
Keep the reasoning behind the next move.
A trading workflow spans research, decisions, execution and review. Memory connects those stages so a fresh session can start with the strategy and constraints already in view.
EXAMPLE WORKFLOWTrading agents
USER REQUEST
Start with the question.
Review this new market signal against our existing strategy.
Start with the question.WHAT THE WORKFLOW CAN REMEMBER01 02 03
Risk preferences
Keep the user’s exposure limits and approval requirements.
Previous review
Recall why a similar signal was rejected last session.
Example inputs and responses for this workflow.
THE PROBLEMWhen the context
When the context
doesn’t carry forward.
The strategy gets lost
A restarted agent sees a signal without the assumptions that originally made it relevant.
Constraints get repeated
Risk preferences and review requirements have to be supplied again in every session.
Decisions lose their context
An execution record alone does not explain which research, assumptions or instructions informed it.
WHAT MEMORY MAKES POSSIBLEKeep what matters.
Keep what matters.
Build on what came before.
01STRATEGY MEMORY
A strategy that survives the session
Persist the thesis, supporting research and invalidation conditions as tagged records. The next session can retrieve the original context before proposing an action.
Thesis
Entry conditions
Invalidation criteria
02TARGETED RECALL
Recall with a purpose
Query related observations through semantic, keyword and tag-based recall. Let the application choose relevant records instead of copying the whole history into a prompt.
Query
Strategy tags
Related observations
03DECISION HISTORY
A record of each decision
Store the inputs, rationale and outcome as separate records. Content-addressed batches give reviewers a reference for checking that retrieved content matches what was written.
Observation
Rationale
Outcome
04WORKFLOW CONTINUITY
Context across the workflow
Use consistent application and agent namespaces to organise research and review. Decide explicitly which records each stage can access.
Research agent
Review agent
Execution notes
THE WORKFLOWFrom a new observation
From a new observation
to lasting context.
Capture the observation
Write the signal, source reference, timestamp and strategy tags.
Recall before reviewing
Retrieve related research, constraints and prior decisions for the application to evaluate.
Record what happened
Persist the chosen action or rejection, its rationale and the eventual outcome.
Memory supports your application’s decision process. A content identifier proves the integrity of a record, not the quality of a trade.
DESIGN YOUR INTEGRATIONA memory layer.
A memory layer.
Your application’s logic.
Connect the memory runtime to the moments your application needs to write, recall and review context.
Explore memory documentation ↗01
Define a namespace for each strategy or account context.
02
Tag records by instrument, strategy and event type.
03
Keep execution permissions and risk checks in the trading application.
04
Keep source references and timestamps alongside generated summaries.
COMMON QUESTIONS
Before you build.
Does Lighthouse execute trades?
This use case describes memory infrastructure. Signal generation, risk validation, approvals and order execution belong to the trading application.
Does memory guarantee better returns?
No. Memory can preserve relevant context and make decisions easier to review. It does not establish that a strategy or market observation is correct.
What should the agent remember?
Useful examples include the original thesis, risk preferences, source references, decisions and outcomes. Your application chooses which records to write and recall.
What can a reviewer verify?
A content identifier can help verify that retrieved bytes match the recorded batch. It does not prove the accuracy of the research or the quality of the decision.