OneMemory Join the waitlist

Every AI you use.
One memory.

OneMemory gives Claude, ChatGPT, Cursor, and your agents a shared long-term memory that lives on your machine.

Join the waitlist How it works
What your AI tools see
claude-code > memory_recall("deploy process for project:atlas") 3 memories returned, scoped to project:atlas   codex > memory_propose({ claim: "Staging deploys run from the release branch", scope: "project:atlas", evidence: { source_type: "chat", excerpt: "..." } }) policy: candidate. Queued for your approval.   you > approve memory active. Audit event recorded.   claude-desktop > memory_recall("deploy process") includes what Codex learned two minutes ago

Your AIs don't talk to each other.

Claude forgets what you told ChatGPT

Each assistant keeps its own siloed memory, if it keeps one at all. You are the only sync layer, and you're tired.

Every project starts from zero

Conventions, decisions, gotchas: re-explained to every tool, every session.

Your context lives in apps you don't control

What your AIs know about you sits in someone else's cloud, on someone else's terms.

Memory with you as the approver.

OneMemory runs quietly on your machine and connects to your AI tools over MCP and REST.

observe

Watches your AI sessions locally

Adapters for Claude Code, Claude Desktop, Codex, and manual saves detect durable facts, decisions, and preferences.

redact

Strips secrets before anything is stored

API keys, tokens, and passwords are removed on the way in, not after the fact.

propose

Turns facts into evidence-backed claims

Every memory carries its source, scope, and confidence. Claims, not gospel.

approve

You decide what becomes memory

Sensitive items wait in a review queue. Nothing important is saved behind your back.

recall

Every AI starts where the last one left off

Scoped retrieval means project A's context never leaks into client B's work.

Your memories never leave your machine.

Local-first storageMemories live in a database on your disk. Cloud backup is optional and off by default.
Consent-gated observerSession watching is off until you turn it on, per client, with a plain-language disclosure.
Hash-chained audit logEvery read, write, approval, and deletion is recorded in a tamper-evident chain.
Full export, alwaysYour data is yours. Export everything to open formats at any time, subscriber or not.

How it compares.

Built-in memories work well inside their own app. OneMemory is for people who work across all of them.

OneMemory ChatGPT memory Claude memory OpenMemory by mem0
One memory across your AI tools YesMCP and REST, any client ChatGPT only Claude apps only YesMCP clients
Lives on your machine Yeslocal database, works offline No, OpenAI account No, Anthropic account Yes
Shared across your computers Yesopt-in sync, host-aware memories Yes, in their cloud Yes, in their cloud Not built in
Watches your AI sessions in the background Yeslocal observer with per-client adapters Only inside its own chats Only inside its own chats No, tools must be called
Decides what's worth keeping Yeslocal policy engine, no API key needed Automatic, server-side Automatic, server-side Via cloud LLM key
You approve what's saved Yesreview queue for sensitive items Auto-saves Auto-saves Auto-saves
Detects contradictions Yesshows you both sides, you resolve No No Auto-merges
Secrets stripped before storage Yesredaction on the way in Not documented Not documented Not documented
Scoped containers per project and client Yesenforced isolation, no cross-leakage Basic projects Basic projects Per-app controls
Evidence and tamper-evident audit Yessource, scope, confidence, hash chain No No Access logs
Setup One installer Built in Built in Docker, Postgres, Qdrant

Reflects each product's publicly documented behavior as of July 2026. They evolve; so do we.

One installer. A whole memory system.

The table is the short version. This is what actually ships, and every part of it runs on your machine.

capture

Observer daemon. The always-on watcher that reads your AI session activity locally, so nothing worth keeping slips by.
Per-client adapters. Purpose-built readers for Claude Code, Claude Desktop, and Codex, plus manual saves and REST ingest.
Secret redaction. Keys, tokens, and passwords stripped before anything touches disk.
Event store. Raw activity queues as events first. Nothing becomes memory by accident.

decide

Intelligent layer. Extracts durable facts from observed activity and scores each one before it goes anywhere.
Policy engine. Routes every candidate: ignore, queue for review, or auto-accept only when explicitly low-risk.
Conflict detection. New claims are checked against what's already stored. Contradictions surface instead of overwriting.
Temporal reasoning. Facts age. Stale claims expire, get invalidated, or get superseded on schedule.

govern

Approval queue. Sensitive items wait for you, with a tray notification when something needs a decision.
Scoped containers. Memories live in project, client, repo, and session containers with enforced isolation between them.
Supersession chains. Corrections link old to new. History is preserved, never silently rewritten.
Hash-chained audit. Every read, write, approval, and deletion in a tamper-evident log.
Per-client keys. Each AI tool gets its own credentials, permissions, and rotation.
Cross-machine sync. Your memories follow you to every computer you work on. Opt-in, and host-aware: facts about one machine stay on that machine.

recall

Hybrid retrieval. Full-text plus local vector embeddings plus reranking, tuned for small precise answers.
Context broker. Builds a compact brief for the task at hand instead of dumping raw memories into your prompt.
Entity graph. People, projects, and tools link into communities so related memories travel together.
Project packs. Portable context bundles a new AI session can load in one call.
Admin dashboard. See what every client read, wrote, and proposed, and step in when you want.

Be first in line.

Private beta opens on Windows first, macOS after. Waitlist members get in before anyone else.

One email when the beta opens. Nothing else, ever.

You're on the list. We'll email you when the beta opens.