Memory & Personalization

🟢 Beginner friendly — no setup beyond chatting. 🔒 Privacy: everything stays in SQLite on your machine.

Minara Agent builds a private memory of how you trade, what you care about, and what you've already told it. Every new conversation picks up where the last one left off. Nothing leaves your device unless you ask it to.

What you can do

• State preferences in plain language — "I never trade meme coins", "reply concisely", "prefer on-chain data over K-line charts". The agent remembers and applies them across future sessions.
• Build a trader profile from real activity — the agent's picture of you blends three sources: trades you executed in chat, perps fills across every Minara sub-wallet (web + mobile included), and spot swaps you did outside the agent. So your picture stays current whether you traded through chat or directly on Minara.
• Persist trade notes — "BTC looked frothy at 70k but pulled back to 62k on low volume" gets stored against the symbol and surfaces next time BTC comes up.
• Customize the system prompt — append a personal prefix ("treat me as a professional trader, show percentages to 2 decimals") that rides on every turn.
• Structured profile tags — Risk Profile: Conservative, Web3 Knowledge Level: Advanced — shape how the agent frames answers.

Manage it in the web UI

The web UI surfaces five Memory pages under the Memory menu. Each one is a thin client of the gateway endpoints (/v1/profile, /v1/memory, etc) so the same data shows up in chat, REPL, and the UI.

• Financial Profile — the LLM-rebuilt summary at the top, then a per-symbol breakdown of your perps activity over the last 90 days, plus your reference wallet watchlist and custom instructions. A "Refresh" button forces a rebuild (throttled to once a minute) and also pulls the latest perps + spot history from Minara before the agent re-summarizes.
• Personalization — the behavioural tags (Finance Knowledge, Markets, Risk Profile, FOMO Index, etc) shown as segmented chips. Tap "Edit" to pick a different value; the source badge ("You", "Inferred", "Learned") tells you whether the agent guessed it or you set it. Below that, learned preferences pending your approval.
• Long-term memory — your written facts grouped by type (Constraints, Preferences, Goals, Observations). The "+ Add memory" button writes a new fact (anything ≤ 500 characters); edits and deletes work only on memories you authored. Deletes are soft — a "Undo" toast appears for 5 seconds, and the row stays recoverable for 30 days through the CLI.
• Trading cases — read-only audit dashboard for the agent's learning loop. Snapshot stats, a 12-week alpha trend chart per methodology, and the top patterns the agent has been using. This page is not a trading recommendation — it's the agent's record of what frameworks have worked for you historically. The data is managed by the agent's reflection cron; the UI never offers edit or delete buttons here.
• Behavior Memory — a private financial-behavior fact layer built from meaningful product actions rather than clickstream prose. Overview shows repeated financial patterns and evidence-backed workflow or feature suggestions; Activity exposes the privacy-safe event timeline; Storage Health monitors the isolated SQLite database and provides cleanup/reset controls. Suggestions never change order parameters or safety gates.

How Behavior Memory combines with Trading Memory

Behavior Memory keeps raw actions in a separate local database. A versioned fusion evaluator reads aggregate behavior features together with read-only summaries of eligible manual Trading Memory decisions. Autopilot, Strategy Studio automatic execution, and unknown-origin trades are excluded from the manual-trading side of this fusion. The result can suggest a missing research or review step, or recommend a relevant feature, but it cannot change Trading Memory scores, position limits, leverage, or execution settings.

Events store compact dimensions such as symbol, asset class, lifecycle stage, result, and amount/risk bands. Prompts, descriptions, wallet/account IDs, exact amounts, balances, positions, and P&L are not stored in behavior events. Visits and dwell time in Memory, Memory Settings, and Settings are excluded because they are administration rather than financial-product behavior. An inferred profile observation becomes long-term memory only after the user confirms it, and recommendation feedback applies a cooldown to avoid repeats.

Behavior Overview and Activity semantics

Overview counts meaningful business usage rather than every UI event:

• Opening a business module counts as one direct interaction.
• Active page time contributes to effective use time without adding another interaction.
• Primary and secondary navigation events do not duplicate page-open counts.
• Memory and Settings administration never contributes to interactions, effective time, feature rankings, profile observations, or recommendations.

Activity keeps the underlying action ID but renders a localized, module-aware description. Generic events appear as Opened Chat or Actively used Markets; registered semantic events can appear as Started a strategy backtest, Viewed an Autopilot strategy, or Executed a manual trade. Older records are never assigned details they did not originally capture. The Module column omits the redundant Navigation suffix while retaining meaningful submodules such as Autopilot · Setup.

Safety on edits

Destructive actions go through one of three confirmation tiers:

• Soft — a Modal with a danger button, then a 5-second "Undo" toast on the success path. Used for single-row tag resets, single memory deletes, single reference wallet removes.
• Hard — a Modal with a "Type DELETE to confirm" text field, then a 5-second undo toast. Used for clearing your custom instructions and bulk-deleting more than 5 memories.
• CLI only — wiping the entire profile is intentionally not exposed in the UI. Use minara profile reset from the terminal.

Example prompts

remember that I only trade BTC, ETH, and stablecoins
set my risk profile to conservative
what do you remember about me?
clear my custom prompt
show my preferences
forget that I asked about SOL last week

Example output

Prompt: what do you remember about me?

Agent response:

Here's your current profile:

Preferences (active)

• Never trade meme coins
• Reply concisely

Profile tags

• Risk Profile: Conservative
• Web3 Knowledge Level: Intermediate
• Preferred Chains: Ethereum, Base, Solana

Recent trade notes

• BTC (2026-04-10): "dipped to 62k on low weekend volume"
• SOL (2026-04-08): "accumulation phase, watch $140 resistance"

You can update any of these with /profile set <tag> <value> or just tell me in plain language.

Commands

• REPL: /profile, /profile set <tag> <value>, /profile prompt <text>, /preferences list, /preferences pending
• CLI: minara profile, minara preferences

Privacy

Everything lives in SQLite under ~/.minara/minara.db on your machine. No cloud sync, no telemetry. You can export or delete it at any time.

How it's built

This feature spans four interlocking stores under System Design's Memory subsection — session memory, personalization, role reflection, and the learning system. For the user-facing "how it gets smarter over time" angle, see Self-Improving Agent. Account activity used for personalization is broader than the manual-only learning corpus; see Trading Memory for that boundary.