Memory & personalization

The vault

Pilot writes to an Obsidian-compatible Markdown vault, scoped per wallet and network.

Pilot/
├── Profile.md        who you are, how you like answers
├── Policy.md         your limits (syncs to policy.set)
├── Journal/          one note per day of validated activity
├── Trades/           one note per validated transaction hash
├── Venues/  Tokens/  research notes
└── Reflections/      a short note after each validated trade

After every validated execution, Pilot writes Pilot/Trades/<hash>.md, appends Pilot/Journal/<date>.md and adds Pilot/Reflections/<hash>.md. Execution metrics it does not have (for example fill price) are written as unavailable, never estimated.

Backends: local filesystem (VAULT_BACKEND=fs) or a GitHub repository (VAULT_BACKEND=github). See Markdown vault.

Structured preferences

With PILOT_STRUCTURED_PREFERENCES=1, an explicitly durable statement like "I prefer concise summaries in future chats" becomes a proposal. You accept, correct, reject or forget it on the Memory page. Nothing is saved silently.

• Categories: communication, presentation, workflow, research.
• Up to 12 accepted preferences are supplied to later chats as quoted, untrusted context.
• "I prefer conservative trade analysis" can reorder which caveats and downside evidence appear first. It never changes an intent, amount, route, limit, approval or execution.
• Repeated facts in your validated-trade journal can produce a pending candidate (for example, familiar assets). One event is never enough, and Pilot does not infer risk tolerance, profitability or venue preference.
• Current-request instructions override remembered style; temporary wording is not persisted.

Workflow lessons

Lessons tune how Pilot works with you: response detail, evidence layout, approval explanations, comparison emphasis and research risk posture. Each goes through propose → your review → approved. All except risk posture need verified experience evidence. Pilot takes the 12 most recent accepted lessons, drops any category the current request overrides, and injects up to 8 into chat.

The hard line

Memory is soft context. It cannot change permissions, limits, amounts, venues, approvals, signing or policy. If you tell Pilot to ignore remembered preferences, personalization is skipped for that request.

Hosted storage

In hosted mode, each preference or lesson document is encrypted before it leaves the app (keys derived from SESSION_SECRET), bounded to 1 MiB, and limited to 4,096 rows per principal. Forget keeps a content-free tombstone for idempotency.