brain — the recorder

brain-mcp — the recorder for your AI conversations

Your AI history is being deleted right now. Claude Code deletes session files older than cleanupPeriodDays (default 30) at startup. Run this and see your own cliff edge:

# macOS

find ~/.claude/projects -name '*.jsonl' -exec stat -f '%Sm %N' -t '%Y-%m-%d' {} + | sort | head -3

Linux

find ~/.claude/projects -name '*.jsonl' -printf '%TY-%Tm-%Td %p\n' | sort | head -3

The oldest date you see is where your history ends. brain-mcp records it before it goes — the whole session, subagent transcripts and saved tool output included, byte-exact, content-hashed, locally — makes it queryable with citations you can verify with sed and shasum, tells you what was deleted, and puts it back.

If Claude Code deleted it

brain-mcp health          # the receipt: "3 session(s) and 41 file(s) are gone from where the

agent kept them; the floor holds every one" — each re-hashed

against the manifest when the recorder first noticed it

brain-mcp restore --list # what can be put back brain-mcp restore # writes the session (and its folder) back, byte-exact, then:

cd && claude --resume

restore never overwrites different bytes (identical bytes are a no-op), is all-or-nothing, and takes --dry-run and --to . It is a CLI verb only — no MCP tool writes outside the floor.

Install

pipx install brain-mcp           # or: uvx brain-mcp

brain-mcp install cc # CC hooks + 60-second scheduler brain-mcp serve # the MCP server (stdio) — add to your client config

Or as a Claude Code plugin (hooks + server in one step):

/plugin marketplace add mordechaipotash/brain-marketplace

/plugin install brain

What it does

• Captures at source. Claude Code via Stop/SessionEnd hooks (line-deltas + a final snapshot); Codex, Pi, and everything inside a Claude Code session's own folder (subagent and workflow transcripts, .meta.json, tool-results/) via a 60-second poll-scan. One spool, one floor.
• Keeps the bytes. The floor is ~/.brain/lake/ / .jsonl — append-only, byte-identical to the original, with an append-only sha256 manifest. Nothing is parsed and discarded; the DuckDB index is a cache, fully re-derivable from the lake.
• Cites or abstains. Every search hit carries {file, line span, sha256} — checkable by anyone with sed -n 'A,Bp' file | shasum -a 256, no database required. Zero hits return an explicit abstention naming the lanes and dates actually searched.
• Reports health honestly. fresh = the floor holds everything the agent's files show; stale = the agent has content the recorder missed; `unknown` = unmeasured, not healthy — never folded into "looks fine".
• Backs up verifiably. brain-mcp backup syncs lake+manifest and re-hashes sampled files at the destination. The verify can fail; that is the point.
• One network fetch, ever, and it is not about you. No telemetry, no cloud, no accounts, and no HTTP client in this package. The exception: DuckDB downloads its full-text-search extension (fts) from DuckDB's own extension repository the first time the index is built, and caches it under ~/.duckdb. brain-mcp install does that fetch up front and says so; brain-mcp doctor reports whether it is cached.

The 7 MCP tools

🤖 For AI Assistants

*You're reading this because your human runs THE RECORDER and gave you access to their AI history. This is the honesty contract.*

The one rule: everything you state from these tools is either CITED or ABSTAINED — there is no third state. A cited claim carries a checkable pointer (file + line span + sha256); verify it with brain_get(expect_sha256=...) before building on it. An abstention means "not found above threshold in the lanes and dates the tool measured" — it does NOT mean "it never happened". Never fill an abstention with your own guess.

• Present recall as their words, dated: *"On 2026-08-19 you wrote: '…' (sess-7f2a.jsonl:412)"* — never as your own knowledge. One claim, one citation.
• verified: false from brain_get means the floor changed since indexing. Say so plainly.
• A health response containing any unknown lane is never "everything looks fine". The honest sentence is: "2 lanes fresh, 1 stale, 1 unmeasured."
• "What do I think about X" → brain_search(query, role="user"). That returns what the PERSON wrote: text the harness injected under the user role (skill bodies, reminders, slash-command wrappers, compaction summaries, Codex context blocks — a quarter to two thirds of "user" rows, measured) is excluded unless you pass include_machine=True, and subagent transcripts unless include_subagents=True. "How did my thinking evolve" → add order="time_asc"; bound it with since / until. The server has no opinion about your human's mind; it has their words, with receipts.

The floor format

~/.brain/

spool/ / hooks + scanner write here (atomic, dot-tmp invisible) lake/ / .jsonl THE FLOOR: append-only, byte-identical to the origin; a rewrite opens .g2.jsonl — old kept, never deleted lake/cc_sessiondir/ / /... a CC session's folder, mirrored at the same relative paths (subagents/, tool-results/, *.meta.json) manifest/manifest.jsonl one versioned line per chunk: byte range, line range, sha256 offsets/ / hook fast-path line counters health/*.last_run side-effect heartbeats (mtimes are the proof, never a report) brain.duckdb the index — a cache, re-derivable from lake/ + manifest/

Where your agents keep their transcripts: Claude Code ~/.claude/projects/**/*.jsonl (rolling window!), Codex ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl, Pi ~/.pi/agent/sessions/**/*.jsonl.

Other verbs

brain-mcp record                  # one capture tick (the scheduler runs this every 60s)

brain-mcp health [--exit-nonzero-on-stale] # cron-able brain-mcp doctor # capture status, health, fts cache, the receipt, runaway generations brain-mcp restore [ ] [--list] [--to DIR] [--dry-run] # put a deleted session back brain-mcp redact --lines A B --reason "..." # tombstone a secret; audited in manifest brain-mcp migrate-v1 # import v1 data (marked v1_derived) brain-mcp uninstall # removes hooks + scheduler; your floor is KEPT

v1 → v2

v2 is a rebuild around one principle: capture the bytes first; derive everything else. v1 parsed conversations into a parquet and discarded the originals — v2's floor makes that structurally impossible. v1's 25 tools became 7: the synthesis tools ("cognitive patterns", "switching cost") are gone because a claim that can't carry a line-span citation isn't one this server makes. Migration: brain-mcp migrate-v1 — v1 rows are kept, marked as derived, and floor-backed rows win wherever the source still exists.

Windows: out of scope for v2.0. Scheduling is LaunchAgent (macOS) / systemd user timer (Linux).

MIT.