brain · git:20261022.867b0e0 · 2026-10-22 · sha256 8eb7a2e84edb426a
brain git:20261022.867b0e0A
Immutable. This exact content never changes and is served at /api/v1/blob/8eb7a2e84edb426a.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_vruvpurn7gzti5fa/8eb7a2e84edb426a6d2b2302d3c40bf2035a77af3574b15325de7b8ebb383e17.json
Then /plugin install brain@markdownregistry-vruvpurn7gzti5fa in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: brain description: "Use when: a long-running session must keep a decision, a delegation window, the current task, the next steps or a trap across compactions — add, close or supersede hot memory items with `brain op`, write the owner's own decisions and windows with `brain owner` (through Claude Code's permission dialog), re-claim the brain after /clear or --fork-session, or read why a brain command was refused." --- # brain — hot memory that survives compaction ## Key rules (read these first; this skill is re-injected after a compaction only up to 5,000 tokens) 1. **Hot memory is re-injected at every `SessionStart`** (startup, resume, clear, compact, fork) as the block `## Brain · hot memory`. Trust that block over your memory of the conversation, and keep it true with item operations. 2. **Only two commands change it.** `brain op …` for Now (`N`), Next (`X`) and Traps (`T`); `brain owner …` for the owner's pinned decisions (`D`) and delegation windows (`W`). Never edit, write or redirect into the state directory (`brain doctor --path`): a guard denies the edit, and a hand change breaks the hash chain and stops every write (rc 5). 3. **Owner words go through `brain owner`, one literal command.** Write the decision text, the owner's quote verbatim and where it was said, with no variable, `$(…)`, backslash, glob or redirect, alone in its Bash call and alone in its turn (any other Bash call voids the pending dialog's record; no `"` in the text or source). The owner then sees one line in Claude Code's permission dialog — `Pin owner decision D3, from "<where>": "<text>"` — and approves or refuses it. No dialog, no write: in `dontAsk`, `plan`, `bypassPermissions`, a `-p` run, a subagent, or while a `PermissionRequest` hook could answer for the owner, it is refused, and you say so to the owner instead of retrying another way. 4. **Text is a delta, never a rewrite.** A changed meaning is `supersede` (both items kept, every anchor carried over, at least 60% of the first text unless `--shorter` with a `--reason`). Done work is `close` with a `--reason` and a `--pointer` (PR, SHA, file). At most 5 retirements per 30 minutes; a deliberate sweep says `--bulk` (logged). 5. **Anchors are the exact strings a summary must keep** (a PR number, a 7-hex SHA, a path, a code word): add them with `--anchor`. The loss score of every compaction counts how many of them the summary dropped. 6. **One writer per project.** After `/clear` or `--fork-session` the session id changes: run `brain claim --take-over` (logged). A subagent never writes: it reports back, and the main session runs `brain op`. ## Commands | Command | Use | |---|---| | `brain op add --kind now\|next\|trap --text T [--anchor A]… [--link L]…` | the current task (≤ 3), next steps (≤ 5), traps (≤ 4) | | `brain op close ID --reason R [--pointer P]` | done or dropped | | `brain op supersede ID --text T --reason R` | the meaning changed | | `brain op touch ID`, `brain op hit ID` | still current; a trap was met again | | `brain owner add --kind decision --text T --owner-quote Q --where W [--anchor A]…` | the owner decided | | `brain owner add --kind window --text T --owner-quote Q --where W --expires 2026-10-17T22:00:00Z` | the owner delegated until a time | | `brain owner unpin ID`, `owner supersede ID --text T`, `owner promote ID POINTER` | the owner retires or replaces one | | `brain render`, `brain doctor` | see the block; check the chain, files, filter, answerers, version | Texts are ≤ 150 characters, sensitive values are refused (masked in the message), and control or invisible characters are refused. At a count cap an add is refused and names the item to close. ## When a command is refused | rc | Meaning | Do | |---|---|---| | 1 | a rule (cap, budget, sensitive value, character, length) | follow the message; do not split or reword around it | | 3 | another session writes this project's brain | `brain claim --take-over` if this is the same work continuing after `/clear` or in a fork | | 4 | a sandbox worker | workers keep no hot memory | | 5 | the log does not match its head | stop writing; run `brain doctor` and tell the owner | | 6 | no approved owner dialog | tell the owner; ask in an interactive session in default, acceptEdits or auto mode | | 7 | no `CLAUDE_CODE_SESSION_ID`: not run inside a Claude Code session | run it from a Claude Code session; do not `claim --take-over` | ## Around compaction - `PreCompact` logs the trigger and the context size; `PostCompact` saves the summary (never injected) and its loss score in the Episode Log (`episodes.jsonl` in the state directory). - `BRAIN_OFF=1` silences every hook (the probe's control arm). - The `# Compact instructions` block for a project CLAUDE.md is `templates/compact-instructions.md` of this plugin.