Blog
How to Make AI Remember Your Portfolio
To make AI remember your portfolio, you need to provide more than just ticker symbols and quantities — you also need to include why you hold each position, what you're trying to verify, what risks you can tolerate, and what changes would overturn your original thesis. nashnova can store this information as long-term memory and update it based on subsequent conversations; users should periodically review and edit or delete outdated content on the long-term memory page.
Holdings are "what you have," investment rationale is "why you have it," and risk preference determines "under what circumstances you would change." Only when all three layers of information are combined can AI connect new earnings reports, industry data, or macro changes to the user's existing judgments.
Key Takeaways
Saving only ticker symbols is usually insufficient to support ongoing research; you also need investment rationale, risks, time horizon, and research objectives.
nashnova can update long-term memory based on conversations, but automatic updates do not mean your holdings are always current.
Users can view, edit, and delete related content on the long-term memory page.
Once a memory item is deleted, it is no longer used as background for subsequent analysis; data processing and retention rules are subject to the official privacy policy.
Long-term memory is used for research context and will not place orders or make buy/sell decisions on behalf of the user.
Why Isn't Giving AI a List of Ticker Symbols Enough?
Suppose two people both hold the same semiconductor stock. One focuses on advanced packaging demand over the next three years, while the other is only doing event-driven research around earnings. The ticker is the same, but the metrics to watch, time horizon, and risks are completely different.
If AI only knows the stock name and quantity, it tends to produce generic company overviews, recent news, or price movement explanations. To determine whether a piece of new information truly matters to the user, AI also needs to know:
How much weight this position carries in the portfolio.
What the original reason for holding was.
What facts support or weaken that reason.
Whether the user's focus is weeks, quarters, or years.
What level of volatility and concentration the user can tolerate.
What changes would prompt the user to re-examine the original thesis.
Complete context won't make AI automatically reach the right conclusion, but it can reduce situations where "it answered a lot of information but didn't answer your question."
What Investment Context Should AI Remember?
A practical portfolio memory includes at least five layers of information: holding facts, investment rationale, risk preferences, monitoring signals, and time constraints.
Information Layer | Example | How AI Can Use It | User Still Needs to Verify |
Holding Facts | Ticker, quantity, portfolio weight, cost range | Identify portfolio exposure, concentration, and related risks | Whether holdings are up to date |
Investment Rationale | Why you hold it, key operating assumptions | Connect new evidence to original thesis | Whether the investment thesis has changed |
Risk Preferences | Tolerable drawdown, exclusion list, concentration limits | Adjust research focus and risk alerts | Whether it still reflects current circumstances |
Monitoring Signals | Earnings metrics, industry data, policy variables | Organize follow-up verification checklists | Whether signals truly affect conclusions |
Time & Constraints | Data cutoff date, research timeline, known data gaps | Avoid presenting outdated information as current facts | Source, timing, and completeness |
Position sizes and costs don't necessarily need to be precise down to every trade. Users can provide portfolio weights or approximate cost ranges based on research needs. More importantly, specify the data cutoff date to prevent AI from treating old holdings as current ones.
A Copy-Ready Portfolio Context Template
When establishing memory for the first time, you can provide information using the following structure:
Asset / Ticker:
Position Size or Portfolio Weight:
Approximate Cost Range (optional):
Reason for Holding / Investment Rationale:
Key Metrics to Verify:
Risks That Could Overturn the Thesis:
Research Time Horizon:
Risk Preferences & Constraints:
Tasks You Want AI to Help With:
Information As-of Date:
For example:
Asset / Ticker: XYZ
Portfolio Weight: ~8%
Reason for Holding: Expecting data center demand to drive core business growth over the next two years
Key Metrics: Data center revenue, gross margin, major customer capex
Risks That Could Overturn the Thesis: Noticeable slowdown in order growth, or gross margin consistently below expectations
Research Time Horizon: Next 4 quarters
Risk Preferences: Don't want any single theme to exceed 25% of the portfolio
Tasks for AI: Check whether new earnings support the original investment thesis and list counter-evidence
Information As-of Date: 2026-08-19
This input doesn't ask AI to judge "should I buy or not" — it tells AI what to check. Subsequent conversations can also more easily continue around the same set of assumptions.
What Difference Does Vague Input vs. Full Context Make?
Input Type | Example | Typical Result |
Ticker only | "I hold XYZ, analyze it for me." | Tends to generate a generic company overview and recent news summary |
Ticker plus position | "XYZ is 8% of my portfolio." | Can discuss concentration, but doesn't know why the user holds it |
Holdings plus rationale | "XYZ is 8%, I'm focused on the data center business." | Can connect new evidence to the business assumption |
Full context | Adding key metrics, risks, time horizon, and portfolio constraints | Easier to distinguish supporting evidence, counter-evidence, and follow-up items |
Longer context isn't necessarily better. Repeated news, irrelevant personal information, and invalidated assumptions add noise. What truly needs to be saved is the background that will still be used in subsequent research.
How Does nashnova Update Long-Term Memory?
nashnova can remember investment rationale, holdings, risk preferences, and related investment context that users save or mention in AI Agent conversations. The system can update this information based on subsequent conversations and use it as analytical background in future research.
For example, if a user explains in a conversation that they've sold a position, or that the original investment thesis has been invalidated due to earnings changes, the relevant context can be updated through the conversation. This way, the next research session doesn't require explaining the entire background from scratch.
However, automatic updates don't mean the system can automatically know about every trade the user completes outside the product. Unless the user supplements changes in conversations or long-term memory, saved information may already be outdated. After every significant position adjustment, change in investment rationale, or shift in risk preferences, you should review your long-term memory.
For more on how Agents use research frameworks and context, read What Is an Investing Agent.
How to View, Edit, or Delete Portfolio Memory?
Users can view the investment rationale, holdings, risk preferences, and other investment context saved by the system on the long-term memory page.
When you find inaccurate information, you should edit it directly rather than relying on subsequent responses to self-correct. Common situations that require updates include:
After completing a buy, sell, or significant position adjustment.
After the original investment thesis is weakened or overturned by new evidence.
After the research time horizon changes.
After tolerable drawdown, concentration limits, or other risk preferences change.
When the information in memory no longer has value for future research.
Users can also delete unwanted memories. Once a memory item is deleted, it is no longer used as background for future analysis. This statement does not mean it is immediately and permanently erased from all technical logs, backups, or copies; specific data processing and retention rules are subject to nashnova's official privacy policy.
What Information Is Appropriate to Save, and What Should Not Be Entered?
The goal of portfolio memory is to support research, not to collect as much personal information as possible. Only provide the minimum information needed to complete the research.
Appropriate as Research Context | Should Not Be Entered for Investment Research |
Holdings, portfolio weights, cost ranges | Brokerage passwords, payment passwords |
Investment rationale and monitoring metrics | SMS verification codes, dynamic tokens |
Risk preferences and research objectives | Full bank card numbers or unrelated identification |
User-defined investment constraints | Private sensitive information unrelated to research |
Whether to provide exact position sizes and costs should be decided by the user based on research needs. If portfolio weights and cost ranges are sufficient, there's no need to enter more details.
How Often Should You Check Memory After Portfolio Changes?
There's no fixed frequency that suits everyone. Rather than checking mechanically every day, a more practical approach is event-driven maintenance:
After completing a buy, sell, or significant rebalancing.
After a company releases important earnings, guidance, or regulatory filings.
After a key assumption in the original investment thesis changes.
After the portfolio's thematic exposure or risk preferences change.
Before asking AI to conduct new portfolio research.
If holdings haven't changed for a long time, you should still verify the information cutoff date before using it again. No matter how thoroughly old data is documented, it doesn't equal current context.
How to Confirm AI Is Using the Correct Portfolio?
Before asking AI to do in-depth research, you can first ask it to recite the current context:
Please first list the holdings, portfolio weights, investment rationale, risk preferences, and information as-of date you currently remember. Flag any items that are uncertain or potentially outdated separately, and do not start the analysis directly.
When verifying, check at least four items:
Are the holdings complete? Are any newly added assets missing, or are sold assets still retained?
Are the weights and costs still reasonable? They don't need to be precise to real-time prices, but clearly outdated data shouldn't be used.
Is the investment rationale still the current view? The user may have already changed their opinion, but the old reasoning may still be in memory.
Is the timing clear? AI should state the information cutoff date rather than assuming all content is up to date.
After confirming the context, then ask questions about portfolio risk, earnings impact, or new evidence verification. This way, you can troubleshoot "memory errors" and "analysis errors" separately.