Operational memory for AI teams

Build without starting over.

Name: Engrams
Author: Engrams

Engrams gives your AI agents shared memory across Claude, ChatGPT, Codex, Cursor, and Gemini. Decisions, tasks, handoffs, and pull-based sync — so every session picks up where the last one left off.

Start 14-day trial See how it works →

npx engrams-init Set up in 60 seconds — detects your tools, writes MCP config, verifies connection.

Scroll

The problem

Every platform builds a silo.

Claude’s memory works in Claude. ChatGPT’s memory works in ChatGPT. Switch tools — start over. Open a new session tomorrow — explain again. Run multiple agents in parallel — none of them share state.

This isn’t a feature gap that gets closed. Anthropic won’t build memory that works inside ChatGPT. OpenAI won’t build memory that works inside Claude. It’s a structural line no platform will cross.

Engrams sits outside the platforms. One memory layer, accessible from any tool via REST API or MCP. You own the data. You control the schema. Your agents read and write to the same source of truth.

How it works

Connect. Store. Continue.

Connect via MCP or API

Add Engrams as an MCP server in Claude, ChatGPT, Codex, Cursor, or Lovable — or call the REST API directly from any agent you build.

Structure your memory

Save decisions, context, learnings, tasks, and episodes with typed primitives. Your AI reads and writes structured data — not free text blobs.

Continue from any tool

Switch services, spawn new agents, take weeks off. Every tool reads the same memory. Your work persists across sessions, platforms, and teams.

See it in action

Cross-AI continuity in three steps.

Save a decision in ChatGPT

remember({type: "decision", content: "Use Postgres for queue — Redis adds ops burden"})

Start Claude Code, run pull

pull({agent: "CC-eng"}) — retrieves decisions, tasks, and context written by other agents since your last session.

Continue without copy-paste

Claude already knows the decision, the rationale, and the current task status. No re-explaining. No context window wasted on catch-up.

Comparison

What native memory doesn’t solve.

Claude’s memory works in Claude. ChatGPT’s memory works in ChatGPT. Mem0 and Zep require code. Engrams works across all of them — no SDK, no code, just a URL.

See full comparison →

What you get

Everything your agents need to stay in sync.

Typed memory

7 memory types — decisions, context, learnings, episodes, tasks, profiles, instructions — each with metadata, embeddings, and semantic search.

Cross-platform

Claude, ChatGPT, Codex, Cursor, Lovable, Gemini, or any HTTP client. Same data, same API key, same source of truth.

Shared projects

Invite collaborators to a shared memory space. Both of your AI tools read and write to the same project — across platforms.

Tasks with versioning

Structured tasks with owner, priority, and version history. Assign work across agents. Track progress without spreadsheets.

Agent orchestration

Register agents, pull updates with pull({agent: "CC-eng"}), and check inboxes for cross-agent handoffs. One memory layer coordinates Claude, ChatGPT, Codex, and Gemini.

Decision analysis

analyze_decision pulls relevant past decisions, learnings, and context to help your AI give you better recommendations — grounded in your own history.

Instructions

Project-level rules and protocols that load first, every session. Define how your AI should behave — format rules, workflows, constraints — and they persist across tools.

FAQ

How is this different from native memory?

Claude Projects has memory now. Why do I need Engrams?
Claude Projects is excellent — inside Claude. Engrams is for when you work in Claude AND ChatGPT AND Codex AND Cursor AND your own agents, and want all of them reading the same memory. Native memory can’t cross platform boundaries. Engrams can.

What about MemoryPlugin or browser extensions?
Those are built for consumers who want auto-capture across chat interfaces. Engrams is built for developers who want typed primitives, API access, multi-agent coordination, and shared projects. Different tools for different needs.

Will platforms just build this themselves?
They’ll build better memory inside their own walls. They won’t build memory that works inside a competitor’s product — that’s a structural incentive problem, not a technical one. The more platforms improve, the more you use several, the more you need a neutral layer connecting them.

Is this hard to set up?
You need an API key and a config block in your MCP settings or Custom GPT Action. If you’re comfortable with API keys, you’ll be running in minutes. This is a developer tool — setup is intentional, not accidental.

Background

Built by someone who got burned.

“I run 4 AI agents across 3 platforms on the same project. Before Engrams, each session started from zero. Now they share decisions, tasks, and context in real time. The system remembers what I’ve decided — even when I don’t.”

— Gustav Käll, founder & daily user

Who it’s for

Built for developers and AI builders.

You run multiple AI agents that need shared state — not isolated conversations.
You build with Claude Code, Codex, Cursor, or custom agents and want persistent memory across sessions.
You coordinate work between Claude, ChatGPT, and Codex and need one source of truth.
You want your AI to make decisions grounded in your project’s actual history — not hallucinated context.
You own your API keys, your data, your stack. Setup is part of the value — not a hurdle.

Pricing

Try for 14 days. Pay when you know.

Trial — 14 days

Free

Full API + MCP access
All 7 memory types
Unlimited projects
Shared projects
No card required

Pro

Coming soon

Everything in trial
Unlimited memories
Priority support
Higher rate limits

Team

Contact

Shared projects across team
Multi-agent coordination
Priority support