Cognitive sovereignty · a practical flow

Keep your memory yours.

loci gives you a memory you own: plain text you can read, move, and delete. The risk is the tool you point at it keeping its own memory of you, server-side, learned across chats, that you cannot see or erase. This flow closes that gap, so your palace stays the only memory of record. Two memories of you is one too many, and only one is yours.

Leans on: templates/CLAUDE-master.md (the gate and the memory rules) and templates/.lociignore.

01 · Make your palace the only memory

If your tool offers its own memory that learns you across sessions, turn it off. Your palace already remembers, on your terms. Nothing about you should accumulate in a place you cannot open. One memory of record, and it is the plain text on your disk.


02 · Tier what you expose

Everything in an always-loaded file is sent every session. Keep the sensitive facts out of the always-on map and into rooms that load only when that work is open. Load-on-demand is a privacy setting, not only a token one: the context you never load is context that never leaves.


03 · Keep the outward gate

The gate you set for shipping code is the same gate for what your companion transmits about you: no names, no secrets, no private shape of your thinking leaves without your yes. Scrub before anything reaches a vendor endpoint or a public surface. The gate is one habit doing two jobs.


04 · Own the deletion

Because your palace is your backup, the tool's copy is disposable: wipe it whenever you like and lose nothing that matters. One caution. Some vendor resets are irreversible, so glance at the store first and port anything that lived only there before you clear it.

If your companion is Claude

loci federates to no one and works across every tool. When the tool is Claude, four concrete settings hold the line:

  • Memory (claude.ai, Settings, Memory): turn off Generate memory from chats and Search and reference chats. Your palace is your memory now.
  • Training and retention (claude.ai, data controls): turn off Improve Claude for everyone. This keeps you on the short retention window and out of the training set.
  • Quiet the client (Claude Code, ~/.claude/settings.json under env): DISABLE_TELEMETRY, DISABLE_ERROR_REPORTING, DISABLE_FEEDBACK_COMMAND. These are presence-based, so a value of "0" still enables the disable. Unset the variable to reverse it. Prefer these named flags over the umbrella CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, which also switches off automatic security updates.
  • Sensitive one-offs: use a temporary chat, which is not saved and not learned from.
The floor If a tier of your thinking must never leave at all, point a local model at the same plain-text palace. Same blueprint, no wire.

You are done when

The tool you work through holds no memory of you that you cannot see, move, or delete, and the only durable record of your mind is the plain text you own.

The vendor forgets; the palace remembers.


Next: the practitioner guide · the design primitives · privacy posture