It re-reads the same files, re-derives the decision you made last Tuesday and asks you a question you already answered. Scrollback is not memory. It dies at compaction and it dies again at session end. ns-brain is the memory that survives both: one static binary that hands the next session what this one learned, before the first turn.
One binary. SQLite underneath. No daemon, no account,
one version check a day, and you can turn it off.
════ NS-BRAIN — SESSION BRIEF ════════════════ now Fri 21 Aug, 12:24 · last session 2d ago git main · clean · 4 unpushed ## GUARDRAILS (binding, always shown) #651 [constraint] Feature freeze until the brain landscape is untangled #106 [constraint] The OSS/commercial line, and the price per seat ## HOW THEY WANT TO WORK #650 Commit before making changes. That is standard behaviour, not something to ask about ## DUE NOW — 1 task [ ] #34 ask about the context budget before proposing anything ## THE LAST SESSION HERE LEFT NOTHING BEHIND 2 commits, 0 memories. Whatever it worked out is gone. Store yours before you finish.
Most people's answer to this is a CLAUDE.md, an AGENTS.md or a .cursorrules. It works for about a fortnight. Then it grows, because the only edit a markdown file offers anybody is an append.
We audited ours against the code it describes. 73,769 bytes, 8.9x growth in 23 days, wrong in six places. Two lines contradicted each other inside the same bullet, both loading into every session, neither one winning. Every one of those lines was true on the day it was written.
A missing line is cheap: the model goes and looks. A stale line is expensive, because the model acts on it, in your voice, with your authority behind it and the output looks exactly like the output it produces when it is right.
Each one exists because a plain document cannot do it and each one changes what a session does rather than what it reads.
A SessionStart hook writes the wall clock, the git state, the rules that bind this project and the way you like to be worked with straight into the window. A second hook matches every prompt you type against the store and injects what it finds. Neither one depends on the model deciding to look, which is the whole point: what depends on being remembered will eventually not be remembered.
Wiring the hooks →Retrieval is FTS5, then relevance × decay × a bounded stakes term made of tier, importance, usage and confidence. Bounded because what a record is should break ties, never override what actually matched: measured on 571 real memories, an unbounded version reordered 54% of result pairs and promoted some by twelve places.
The scoring model →"TLS terminates at nginx on 8443" never decays. "The box is rebooting" halves every four days and is gone inside a fortnight with nobody remembering to delete it. The type you pick sets the tier and the tier sets the half-life. Nothing is deleted by decay; it just stops outranking things that still matter.
Types and tiers →Store one with --supersedes and the identity moves to the new fact. Ask in the old wording and you get today's answer, not both at equal weight. Try to file a second memory about the same subject in one scope and the write path refuses, names the memory already there and gives you the three ways out.
A memory about code records the commit it was written against. Recall renders a verdict per fact: tree unchanged since stored, or N commits since, re-check. So a brain does not abolish the re-read. It routes it, one fact at a time and tells you which claims are safe to act on without opening a file.
Provenance and verification →A recall miss returns nothing. Not the nearest thing, not the whole scope. That rule exists because a miss once fed a delete loop and destroyed real memories and the warning that would have stopped it lived only in output nobody was reading. The commands that delete in bulk — prune, wipe, uninstall — are a dry run first.
The last line of that result is the part people miss. The memory knows which commit it was written against, so a later session can see nothing has moved and skip re-deriving the whole thing from the files.
$ ns-brain recall "why postgres here" #31 [channel:design/decision] (score 2.114) Postgres, not SQLite: three writers Settled 2026-03-04. WAL contention was real at 3 concurrent writers, measured at 40ms p99 on the deploy box. Deviates from the house SQLite default. (confirmed · 2026-03-04 · source: bench/wal-3w.log) tree unchanged since stored (a41f22c9)
Scopes nest on : and flow inward only. A recall in channel:deploy:prod also sees channel:deploy, channel and global, never the reverse.
A fact written down once, in whatever repository you happened to be in that day, is there the next time it decides an answer. Nobody has to remember it exists and nobody has to go looking.
$ (in the billing repo) "the old charges API is being deprecated. what breaks for us?" ... the general answer, which is correct and true of everybody ... #412 [payments/decision] (stored by another project) Nothing here breaks. You moved off charges during the invoice rewrite, and this repo has never called it.
That last part is the whole product. The general answer was free and useless. The useful one needed a fact about you that lived in a different repository and it surfaced on its own because it matched the question.
Never the widest possible action. The rule is written into the project's own instructions and every command added since has had to answer three questions: what does it do on a miss, on an empty filter and on a typo.
$ ns-brain remember --type gotcha \ --title "staging deploy needs the vault unsealed" #42 in channel:deploy is already about this (67% title overlap): stored: the staging deploy needs the vault unsealed first yours: staging deploy needs the vault unsealed Storing both would put two answers to one question in the brain. Pick one: the fact CHANGED --supersedes 42 same fact, better words ns-brain update 42 --body ... genuinely separate --allow-duplicate
forget refuses pinned memories and guardrails and refuses to delete a memory that others were superseded by: nulling those links republishes facts retracted on purpose.prune is a dry run, dumps the whole brain before deleting anything, refuses the two tiers that never decay and caps at 25 deletions without --yes.forget --all empties one project on purpose: a dry run, then --apply --i-mean-it, with a project-filtered dump written first.ns-brain keeps its own memory in ns-brain. Every figure here comes out of that store or the git history behind it.
Open Claude Code, Codex, Cursor, Gemini or anything that can fetch a URL and say install https://brain.fightclub.pro/. Everything the agent needs is on this page, in llms.txt and in install.md. Nothing to sign up for and nothing to configure first.
$ curl -fsSL https://brain.fightclub.pro/install.sh | sh # then, in the project you want it to remember $ ns-brain init initialized ./brain/brain.db · SessionStart and UserPromptSubmit hooks wired $ ns-brain bootstrap # only if the project has history
brain.db is the only copy of that project's memory.init reports another memory store already loaded, do not delete it. Move what is still true across and let the user delete the rest. Only they can say which half is still true.Ranking, decay, supersession, provenance, the guards, MCP and full export ship in the free edition exactly as they ship in the paid one. A price multiple should buy a different thing, not the same thing with permissions.
Any number of projects and machines, the whole kernel at full strength, on hardware you control. No account and no telemetry. The only call it makes on its own is a once-a-day check for a newer version, which NS_BRAIN_NO_UPDATE_CHECK=1 turns off. One machine needs nothing but the binary; sharing a brain between machines means one of them holds it.
The brain stays on a machine you own. A laptop on mobile data reaches it through a relay you also own, which sees ciphertext and nothing else. No inbound port, no VPN and free however many machines you add.
Read →We run the server. One brain reachable from every machine, writes that queue when the network is not there, a second person on the same store, roles and attribution and encryption at rest. Priced per company rather than per seat: $119.88 a year for one person, $99 a month up to twenty people and $499 up to two hundred, which is $2.50 a head at the top tier.
What you get →The same code as a licensed appliance on your hardware, with your backups and your keys. The licence verifies locally, so a firewall change on your side can never stop your brain answering.
How it works →Free for one developer, on as many projects and machines as you want, for as long as you want it.
$ curl -fsSL https://brain.fightclub.pro/install.sh | sh $ cd ~/projects/yours && ns-brain init