Skip to content
forgetby Multi-turn
Early access
Hosted ledger now on the 0.6 engine

One memory
for every AI you use.

Tell one AI, and the others remember. Forget connects Grok, ChatGPT, Claude, Muse, Cursor, Claude Code, Codex and any MCP agent to one long-term memory. It keeps what was said word for word, answers with a source on every line, and says so when the record is silent.

Sign-up is invite-only during early access.

forget-ai 0.5.1 · forget-connect 0.6.1By Multi-turn

Don’t rewrite memory.
Rebuild it.

Most memory layers summarize or extract facts when they save. Whatever the summary drops is gone, and the answer loses its source. Forget never rewrites the record. It does the work when a question arrives.

Store-time rewriteTypical memory layers
Conversation→Summarize · extract facts→Similar-text search→Answer
details lostno sourceanswers anyway
Query-time reconstructionForget
Conversation→Verbatim recordtime · speaker · source→Rebuilt for the questionevery line cited→Answer or abstain
nothing lostcites sourcesknows when it doesn’t know

One memory,
three moments.

Forget works when a message is saved, while your agents are idle, and when a question comes in. Every derived item points back to the verbatim record, so any answer can be checked against what was actually said.

KnowledgeEpisodesVerbatim
01 · On save
02 · While idle
03 · On question
Scope & policyPer-user scope, deletion, local-first
People & topic pagesLasting facts, preferences, standing instructions
Memory modelSelect & reconstructEvery sentence cites a raw record
Hybrid indexBM25 + embeddings, episode boundaries
Timelines & value historyOrder, current value, previous values
Search & sufficiency checkRe-search what is missing, abstain without evidence
Verbatim recordAppend-only. Time, speaker and source on every turn.Every derived item points here
Answerwith sources
Abstainwhen evidence is missing

↻ Real usage teaches the memory model what to keep, link and recall.

Decide once.
Every tool knows.

A decision made in Claude Code on Tuesday is there when Codex opens on Wednesday. When your server agent runs that night, the version it replaced is flagged before it can be used.

Claude CodeMacBook · Tue 10:14
> use Postgres for billing, not DynamoDB. we need joins for invoices.• forget.add_memory + saved · billing DB is Postgres• forget.supersede_memory × superseded · billing uses DynamoDB
CodexWorkstation · Wed 09:02
> add the invoices table• forget.search_memories · high an LLM reads 40 candidates… ↳ billing DB is Postgres green · Claude Code · Tue 10:14→ 0042_invoices.sql for Postgres
Your agentServer · Thu 02:00
• forget.prepare_context_autopilot goal: invoices v1 next: backfill 2025 invoices• forget.search_memories ↳ green · billing DB is Postgres ↳ red · billing uses DynamoDB→ backfill runs on Postgres

An example project in a simplified terminal. Every step is a real Forget tool call.

Memory you can check.

Decisions, not transcripts.

An observation gate keeps the facts that matter and lets the chatter go. Decisions also surface from your git history, backdated to the commit.

A receipt on every answer.

Every memory links to the commit or conversation that decided it, with a trust label: green if you said it or a tool saw it, yellow if an agent inferred it, red once it is superseded.

Stale facts get caught.

A changed fact is superseded, not deleted. An agent about to act on the old version sees both, labeled, before it acts.

Every session opens briefed.

Start Claude Code or Codex and a capsule is waiting: the current goal, the next step and anything that has gone stale.

Recall, measured.

How often the right memory lands in the top six results, on our hardest benchmark: 120 questions, six question types, multi-part answers scored by coverage.

Instant
Search only, no LLM
89.2%
Deep
An LLM reads 40 candidates · about 3 s
95.0%
Extra
Reads 100 in full · about 5 s
96.7%
Ceiling
Reads every stored memory
99.2%

Every number was measured before its engine shipped. Earlier end-to-end result (July 2026, an older research pipeline, not the current product engine): 78.4% ± 0.4 on LongMemEval (full 500, three runs), with memory built by a local Qwen 14B model and answers read and graded by GPT-4o.

Measured, with the final result still pending.

On the Agent Memory Leaderboard, an independent benchmark where a fixed reader answers only from what the memory returns.

57.93
AML Cycle 2 Full, provisional score

All 9,602 questions finished; organizers' review ongoing, final board mid-November

72.5
Safety & Privacy, AML Cycle 2 Full (provisional)

Robust refusal, minimal disclosure and evidence boundaries, scored as one capability

71.2
Memory Governance, AML Cycle 2 Full (provisional)

Keeping, updating and retiring memories correctly over time

The 57.93 is a provisional score; the organizers are still reviewing Cycle 2 and the final board comes in mid-November, so it may change. Safety & Privacy and Memory Governance are capability scores from the same provisional Full run.

Connect in a minute.

Two ways in. Connect the AI assistants you already use from your account, or run the engine yourself and connect your developer tools.

01 · AI ASSISTANTS AND TOOLS

Connect from your account

Sign in at cloud.forget.sh/connect and follow the card for each app.

  • GrokAdd the connector URL and sign in.
  • ChatGPT · dotAdd Forget as a custom MCP server. dot uses the same connection.
  • ClaudeAdd Forget as a custom connector.
  • MuseCreate an assistant key, then ask Muse to build a custom connector.*
  • Instinct and othersCreate an assistant key and point the assistant at cloud.forget.sh/agent.
  • Cursor · VS CodeOne-click install.
  • Claude Code · CodexA line or two, shown on the connect page.
02 · SELF-HOSTED ENGINE

Run it yourself

One engine, one command per machine. Using an AI agent? Point it at forget.sh/llms.txt and it can set this up for you.

01

Install the engine

curl -fsSL forget.sh | sh

Installs the Forget engine and connects Claude Code, Codex and Claude Desktop. Your existing MCP setup is backed up first.

02

Share one ledger

FORGET_API_KEY=<token> npx forget-connect --url <ledger-url>

Run it on each machine and every client writes to the same ledger. Host it on your own server, or get early access and we host it and send you its URL and token.

03

Undo in one command

npx forget-connect disconnect

Removes the Forget MCP entry, memory-rules block and hooks it added. It edits the files in place rather than restoring a backup; the backups made at connect time stay next to them. Other settings are left as they were.

GrokChatGPT · dotClaudeMuse*InstinctCursorVS CodeClaude CodeCodexClaude DesktopAny MCP client

* Muse is currently available only in the US. Meta does not review custom connectors. Product names are trademarks of their owners; no affiliation is implied.

Pricing.

Forget Cloud

Free

  • Personal memory, free during early access
  • Deep recall included, with per-account fair-use limits
  • Connect your assistants and developer tools to one memory
  • One isolated store per account, hosted in Seoul
  • Invite-only during early access
Connect your AI →

Teams

Preparing

  • One shared memory for a team’s people and agents
  • Paid plans are being prepared
  • For now, personal memory is free
  • 유료 플랜은 준비 중이에요. 지금은 개인 기억을 무료로 쓰실 수 있어요.
Join early access →

Self-hosted

Free

  • The full engine, on your machine or your own server
  • Every connected client shares one ledger
  • Deep recall on your own local model
View on PyPI ↗

Questions.

What exactly gets shared?

Memories: decisions, facts and task state, each with its source. Every client connected to the same ledger reads and writes it. Which tool wrote a memory is kept as provenance, not as a wall.

Where does my memory live?

Self-hosted, it stays on your machine or your own server. With Forget Cloud (early access), it lives in a store of its own for your account, on AWS in Seoul; the sign-in and connection server runs on Fly.io in the US and keeps no request or response bodies. Details are in the Privacy Policy. End-to-end encrypted sync is in design and not shipped yet.

How is this different from ChatGPT or Claude memory?

Built-in memory stays inside one vendor’s app, and your other assistants can’t use it. Forget plugs into those assistants as well as your developer tools, so ChatGPT, Claude, Grok and Claude Code all read and write the same memory. It keeps the verbatim record, and every answer carries a receipt: the conversation or commit that decided it.

What is query-time reconstruction?

Instead of compressing conversations into facts when they are saved, Forget keeps them word for word and assembles the evidence a question needs when it is asked. Each sentence of that evidence cites the raw record it came from, and if nothing supports an answer, the agent is told to abstain.

What happens when a fact changes?

The old fact is not deleted. It is superseded: it keeps its history, turns red, and the new fact links back to it. If an agent is about to act on the stale version, the conflict alert shows both first.

What does it cost?

Personal use is free. Forget Cloud is free during early access, deep recall included, with per-account fair-use limits, and the engine is free to self-host. Paid plans are being prepared. For now, you can use personal memory for free. 유료 플랜은 준비 중이에요.

How do I remove it?

Run npx forget-connect disconnect to remove the Forget entries it added (MCP server, memory rules, hooks) from your client configs; your other settings are left as they were. Self-hosted, delete the ledger file. On Forget Cloud, delete all of your memory or your whole account from your account page, or write to hello@forget.sh.

What is supported today?

From your account at cloud.forget.sh/connect: Grok, ChatGPT (dot uses the same connection), Claude, Muse (US only), Instinct and other assistants with an assistant key, Cursor, VS Code (Copilot), Claude Code and Codex. Self-hosted on macOS and Linux: Claude Code, Codex and Claude Desktop through forget-connect, and any MCP client over HTTP. Not shipped yet: Windows for the self-hosted engine and end-to-end encrypted sync.

Recently shipped.

  • Oct 19

    Hosted ledger moves to the 0.6 engine.

    Every message is now kept verbatim alongside its facts, search combines BM25 with embeddings, and each result reports whether the evidence is strong, weak or missing.

  • Sep 24

    Shared ledger in production.

    Claude Code, Codex and server agents can now share one hosted ledger: one ledger file per user, hashed per-user tokens, daily backups.

  • Sep 5

    0.5.0, the honesty release.

    An unverified fact with a date or amount now asks for confirmation before use, machine-transcribed facts wait in quarantine, and search weighs sources. Learned forgetting, plus a consolidation pass you can run yourself (not scheduled automatically).

  • Aug 4

    The recall dial.

    Four measured gears on one dial, and deep recall on the certified model.

All releases ↗
Early access

Share one memory across everything.

Invite-only for now. Tell us what should share a memory, and we will send you an invite.