Project Memory

The assistant starts every conversation knowing nothing about your server beyond what it can read from the files. Memory is how that stops being true.

There are two halves, and they are separate on purpose: one you write, one it writes.

BOLTISE.md — the half you write

Create a file called BOLTISE.md at the root of your project, beside server.cfg. It is read on every single request.

This is where the things the assistant cannot infer go: which framework you are on when the files are ambiguous, your naming conventions, folders it must not touch, decisions you already made and do not want relitigated.

# This server

- QBox, with ox_inventory and ox_lib. Not qb-inventory — we migrated off it.
- Custom resources live in `resources/[custom]` and are prefixed `srv_`.
- Never edit anything under `[core]`. It is upstream and gets replaced wholesale.
- Money is handled through `srv_banking` exports, never by touching the
  `players` table directly.

Nothing in Boltise writes to this file, so it stays exactly as you left it. Keep it short — it is read every turn, and a page of rules the assistant must weigh on every request is worth less than five rules it always follows.

Boltise also picks up rules files you may already have from other tools: .rules, .cursorrules, .windsurfrules, .clinerules, CLAUDE.md, GEMINI.md and AGENT.md.

Remembered decisions — the half it writes

The assistant can also save things itself, during a conversation, when you tell it something worth keeping. Those are stored against your account rather than in your project folder, and they sync across machines.

Two kinds behave differently:

Standing rules — anything you pinned, and anything saved as a decision or preference — are loaded on every turn, regardless of what you asked. That is the whole point of a rule: “never change endpoint_add_udp” has to be in front of the assistant on the turn it would otherwise break it, and it will not think to search for a rule it does not know exists.
Background memories are matched against what you actually typed, and only loaded when relevant.

Up to 12 standing rules are carried per turn.

Managing what it remembers

Open Settings ▸ AI Memory to see everything saved, search it, and delete anything wrong. You can also ask the assistant directly — it can list, search and save memories as part of a normal conversation.

Saving a memory is the one write that does not stop to ask for approval. It still appears in the audit trail. See Approvals & Safety.

Scope

Memories are either global to your account or attached to one project. Project memories from one server never appear in another server’s chat — if Boltise cannot positively identify which project it is in, it falls back to global memories only rather than risk leaking one server’s conventions into another.

Docs privacy