switch to 3ngram · under 5 minutes · three starting points

whatever you're using to hold your context today — Claude Code MEMORY.md, ChatGPT / Claude exports, or nothing at all — 3ngram imports what's already there and captures everything from here.

beta · no card · we import, you verify, agents stay off until you flip them on.

path 01 — from Claude Codemost common · ~3 min

we read your MEMORY.md and turn it into typed, searchable context.

Claude Code stores per-project memory at ~/.claude/projects/<slug>/memory/MEMORY.md. Upload a single file or a zip of the whole directory. Every top-level heading becomes a memory candidate you can review before anything is saved.

Claude Code export · .md or .zip · up to 50 MB get started free →
00:00 export MEMORY.md (or zip the memory directory). 00:40 upload in Settings → Import. 02:10 3ngram parses each heading into a memory candidate. 03:00 you review · import · agents stay off by default.
path 02 — from ChatGPT or Claudeconversation export · ~4 min

your flat chat history becomes typed — commitments with deadlines, decisions with rationale.

Export conversations.json from ChatGPT or Claude. 3ngram linearises each conversation and runs an LLM extraction pass that proposes memory candidates typed as preferences, decisions, or commitments. Nothing is written until you approve.

conversations.json · your tenant only · first 20 conversations get started free →
00:00 request export data from ChatGPT or Claude. 00:30 upload conversations.json — 3ngram types each memory. 03:20 review candidates · flag any re-typing. 04:00 nothing lost · everything now has shape.
path 03 — from nothingcold start · ~4 min

you have no shared work state yet, and that's fine. 3ngram starts capturing today.

The fastest path is the one with no import. Connect Claude / ChatGPT / Cursor / Codex via MCP, and the next time you finish a session, you tell the agent "debrief this to 3ngram". Only then does anything get saved. Your first capture lands on the next session you debrief.

MCP server · nothing saved without your say-so start from zero →
00:00 follow connecting clients in the docs (Claude Desktop · Cursor · Codex · Claude Code). 01:00 connect the client you already use. 02:30 finish a session · ask the agent to debrief to 3ngram. 04:00 your first capture is in — exactly what you asked to save.
safety rails on day one · defaults we think are sane
agents are off by default.
capture runs. nothing sends. you flip on autonomous actions one verb at a time (draft, send, book, file).
safe
every connected source is read-only until you say otherwise.
3ngram cannot send from gmail, post in slack, or open a PR without a signed toggle in settings.
safe
nothing is mixed with other tenants' data in queries.
row-level security on every query, enforced at the database layer.
safe
imports are deduplicated by content hash.
imported Claude Code memory or chat export already in 3ngram? it won't land twice.
safe

onboarded in about four minutes. no card, no install.

get started free ↵

prefer the full walkthrough? read the import guide ↗ · connecting clients ↗