01
Remember
Your assistant calls remember when something worth keeping happens: a decision, a commitment, a blocker. Or you capture it yourself in one line.
memory layer · model context protocolYour AI forgets.
Your AI forgets.
3ngram remembers.
A persistent memory layer for your AI tools over MCP. Decisions, commitments, and context are saved explicitly, then available to search from Claude, ChatGPT, Cursor, and Codex in later sessions.
No credit card · Works with Claude, ChatGPT, Cursor & more
recall → context
- decision
Ship the pricing page without the Teams tier until seats land.
- commitment
Send Amara the migration runbook by Thursday.
- blocker
Staging deploys blocked on the expired Postgres cert.
Works inside your MCP clients · one shared context · see setup docs ↗
the problem
Every chat starts from zero. You re-explain the project, re-paste the context, and re-make decisions you already made.
Retrieval is the entry point; follow-through is the product.
01·how it works
Remember. Recall. Carried forward.
No new app to live in. Save the work state once, then ask any connected client to search it or start with a scoped briefing.
02
Recall
On later turns, ask any connected MCP client to search. 3ngram retrieves relevant memories by meaning as well as text.
03
Carried forward
Start a session with a scoped briefing when you want orientation. It returns open commitments, blockers, stale candidates, and recent decisions for the client to use.
02·what you get
Built for work you're accountable for.
A memory layer you can read, audit, and correct. Never a black box that quietly rewrites what your tools know.
Semantic recall
Memories surface by meaning, not keywords. Ask about the migration and get the decision, the blocker, and who owns it.
Scopes
Work, personal, shared: memories carry a namespace. Deleting a scope never touches the memories that carry it.
Advisory consolidation
Background jobs propose merging near-duplicates. Proposals wait for your review; nothing is rewritten behind your back.
Content-free reporting
Operational logs, traces, and metrics carry IDs, types, lengths, hashes, and counts, not memory text. The authenticated dashboard is where you read your content.
Audit trail
Every memory keeps its history: the actor class that captured it, when it changed, and what superseded it. Corrections version; they never destroy.
One MCP endpoint
The same mcp.3ngram.ai/mcp URL in every client, authorized in the browser. No per-tool exports, no copy-paste context.
the product
Everything your AI knows, in one place you can read.
Browse and search every memory, watch open commitments and blockers, and inspect each record's lifecycle. The dashboard shows your records; operational telemetry stays content-free.
Illustrative preview with sample memories.
06 ·open source
Open source. Run the memory backend yourself.
The memory engine and its MCP, REST, SDK, CLI, and self-host surfaces are Apache-2.0. The managed dashboard, billing, and cloud operations are proprietary and optional.
$ git clone https://github.com/B3dmar/3ngram
$ cd 3ngram
$ cp .env.selfhost.example .env.selfhost
$ docker compose --env-file .env.selfhost -f compose.selfhost.yml run --rm migrations
$ docker compose --env-file .env.selfhost \
-f compose.selfhost.yml up -d
✓ 3ngram server listening on :3000Apache-2.0 core
The memory engine, MCP server, REST API, SDK, and CLI are Apache-2.0. Read them, fork them, and run the backend on infrastructure you control.
Self-host the backend
Run the server with Postgres, pgvector, and Redis through Docker Compose. The managed service uses the same public engine behind a proprietary hosted dashboard.
MCP-first, no lock-in
Self-host or use the managed service through the same MCP and REST contracts. The typed SDK and CLI are available from npm.
plans ·2 plans · monthly or annual
Start free. Upgrade when you need more room.
Free includes 2,000 live memories and one active MCP client. Pro is $19 / €19 monthly or $190 / €190 annually, with unlimited live memories and MCP clients under reasonable security, rate, and LLM fair-use limits.
Free includes 2,000 live memories and one active MCP client. Pro removes those plan-level caps, subject to fair use.
For one person connecting a first AI client to persistent, typed memory.
- 2,000 live memories
- 1 active MCP client
- Typed memory capture and recall
- Semantic and full-text search
- Briefing and debrief prompts
- Web dashboard
- Account data export
Start free
No credit card required.
For one person using the same memory across multiple AI clients.
- Everything in Free
- Unlimited live memories*
- Unlimited MCP clients*
- Monthly or annual billing
Get Pro
*Reasonable security, rate, and LLM fair-use limits apply.
feature table ·Free, Pro
What each plan includes
| Feature | Free | Pro |
|---|---|---|
| Live memories | Free2,000 | ProUnlimited* |
| Active MCP clients | Free1 | ProUnlimited* |
| Typed memory capture and recall | FreeIncluded | ProIncluded |
| Semantic and full-text search | FreeIncluded | ProIncluded |
| Briefing and debrief prompts | FreeIncluded | ProIncluded |
| Web dashboard | FreeIncluded | ProIncluded |
| Account data export | FreeIncluded | ProIncluded |
Want the full product surface? Read the use-case guides
Give your AI a memory that carries work forward.
Connect Claude, ChatGPT, Cursor, and Codex to the same memory. Save what matters, then search or brief it when the work continues.