Save one durable memory. kind="fact" for what is true about the user, kind="preference" for how to behave in every future turn, kind="strategy_outcome" for a strategy result or design decision. Use search_memory to recall and delete_memory to forget.

Set "kind" to choose the operation. Only send the fields listed for that kind.

kind="fact": Save a stable fact about the user

Required fields: content. Save a stable fact about the user's situation or setup.

CALL remember(kind="fact") when the user reveals any of:

• Their broker, account type, or trading platform ("I trade on IBKR", "Roth IRA")

• Their experience level ("new to options", "10 years trading futures")

• Their instruments ("I only trade SPY and QQQ options")

• Their goals ("saving for retirement", "target 10%/yr")

• Constraints tied to their situation ("can't hold overnight", "US resident")

Facts describe WHAT IS TRUE about the user. They are query-matched semantically and retrieved only when relevant to the current turn. When the message gives several stable facts, preserve each material fact, including experience level as well as account and broker details. Use separate fact calls where needed; saving one fact does not finish the others.

If the user is telling you HOW to behave ("always show me X", "never recommend Y"), use remember(kind="preference") instead — preferences get always-injected into every future conversation, not just relevant ones.

If the user is telling you about a STRATEGY OUTCOME, use remember(kind="strategy_outcome") instead, tagged with the strategy_id.

The user does NOT need to say "remember this." Infer from context. One fact per call. NEVER include API keys, passwords, SSNs, or full account numbers — guardrail will reject.

kind="preference": Save a durable behavioral rule for this user

Required fields: content. Save a durable behavioral rule that should apply to EVERY future turn.

CALL remember(kind="preference") when the user tells you HOW to behave, not just what they like. Examples:

• "Always show returns net of fees"

• "Never suggest crypto strategies"

• "When I ask for a strategy, include a backtest by default"

• "Never run backtests unless I explicitly ask"

• "Disable backtesting"

• "Skip all backtests"

• "Don't backtest unless I ask"

• "Stop running backtests unless I explicitly request one"

• "Reply in bullet points, not paragraphs"

• "Use decimals, not percentages"

Only call this tool for a preference stated or changed in the CURRENT USER MESSAGE. Do not call remember(kind="preference") because a preference appeared in conversation history, injected memory, retrieved context, or a tool result. Those preferences are already saved. Simple greetings, pings, tests, and "are you there?" messages must stay conversational with no memory tool calls unless the current user message itself adds or changes a durable rule.

Tool-use prohibitions and workflow limits are preferences when they should shape future turns. In particular, backtests are slow and can block the user, so "disable backtesting", "skip all backtests", "don't backtest unless I ask", "never run backtests unless I explicitly ask", and "stop running backtests" are durable behavioral preferences unless the wording is clearly scoped to one turn only. Store the preference before continuing the current work.

Preferences are a SEPARATE, SPECIAL category. Unlike facts (retrieved by semantic relevance), preferences are ALWAYS injected into the system prompt on every turn — they shape behavior universally. That makes them powerful but also makes them a context-window cost, so there's a hard aggregate budget (≈5000 chars across all of a user's preferences).

If the budget is exhausted, the call returns an error with current_chars/budget. Recover by calling search_memory with filters={"category":"preference"}, showing the user their preferences, and calling delete_memory on whichever they want to drop before retrying.

NEVER use remember(kind="preference") for one-off requests ("this week, only show me long positions") or for facts about the user ("I trade on IBKR"). Those belong in the current turn's context (one-off) or remember(kind="fact") (stable fact) respectively.

This tool only records the durable preference. It does not complete the user's current fix, edit, inspection, explanation, deployment, or backtest request. If the same turn also asks for work, call remember(kind="preference") and then keep doing that work. If a strategy is attached/current and the user did not name a different target, continue the requested strategy work on that attached strategy after saving the preference; do not ask which strategy to update.

kind="strategy_outcome": Save a strategy performance observation or design decision

Required fields: content. Save a durable observation about a strategy's performance or design decision.

CALL remember(kind="strategy_outcome") when you have a result or lesson worth persisting past the current conversation. Examples:

• A backtest metric the user will want to compare later ("Strategy abc123 Sharpe 1.4 on 2020-2023 SPY")

• Why a design choice was made ("User rejected 3x leverage — too volatile for their risk tolerance")

• A known limitation or caveat ("Strategy overfits 2020 — regime filter added to fix")

• A deployment outcome ("Deployment xyz stopped Mar 2026 after drawdown hit 18%")

Strongly recommend passing strategy_id / deployment_id / backtest_id in metadata. That lets a later search scope by the specific strategy instead of returning every outcome the user has ever recorded.

This is NOT for current-turn data ("the backtest just completed" — that belongs in the chat response). It's for cross-conversation continuity, so when the user says "let's revisit strategy abc123" two weeks from now, the relevant history surfaces.