Engram Blog
How we build Engram, what we've learned, and what we'd do differently. Mostly technical, occasionally opinionated.
· 7 min read · backup, export, import, chatgpt, claude, portability, memory
You back up your photos and your files. The thousands of hours you've spent thinking with ChatGPT and Claude live in someone else's app, with no copy anywhere. Export them, upload once to Engram, and you get a backup that never expires — and that every AI you use can search.
· 5 min read · students, pricing, memory, chatgpt, claude, research
Everything in Engram Pro — 1,000,000 messages of permanent memory, in every AI tool you use — for $3 a month, if you can verify a university email address. What's included, how the one-time verification works, and the honest caveat: we check that you control a university mailbox, not that you're enrolled this semester.
· 7 min read · privacy, consulting, confidentiality, exclusions, cli, compliance
If you bill more than one client, AI memory quietly crosses a line your contracts draw. Three controls that keep each client's work inside its own boundary — and why the one that feels most like control is the one you shouldn't build on.
· 9 min read · comparison, claude, memory, verbatim, mcp, portability
Claude's memory is genuinely good — and it stores short topic files it wrote about you, only inside Claude. Engram stores the conversations themselves, verbatim, across twelve tools. An honest comparison, including where Claude wins.
· 8 min read · privacy, deletion, exclusions, cli, control, gdpr
A memory you can't control isn't yours. How to keep whole repos and individual sessions out of Engram, say "don't save this" mid-conversation, and delete what did get in — one message, one conversation, or everything.
· 7 min read · portability, lock-in, chatgpt, claude, import, migration
The real cost of changing assistants isn't learning a new one — it's the years of context you'd leave behind. Both ChatGPT and Claude will hand you that history today. Here's how to make it searchable from whatever you use next.
· 6 min read · mcp, memory, integrations, announcement, architecture
Engram now works in twelve AI tools — ChatGPT, Claude, Claude Code, Cursor, VS Code, Windsurf, Zed, Cline, Goose, Gemini CLI, Codex CLI, and JetBrains. We didn't write twelve integrations. That's the point, and it's why this list only grows.
· 7 min read · comparison, memoryplugin, mcp, memory, chatgpt, claude-code
Both give your AI a memory that follows you between tools — with opposite architectures. A fair comparison: extension vs MCP, distilled notes vs verbatim transcripts, where each one wins, and how to pick.
· 7 min read · claude-code, context-compaction, memory, mcp, how-to
When Claude Code compacts, it replaces your conversation with a lossy summary — and the agent trusts it, confidently rebuilding decisions it half-remembers wrong. Here's how to recover the exact words, and the CLAUDE.md rule that makes it automatic.
· 4 min read · chatgpt, cli, how-to, conversation-full, import
A full ChatGPT conversation can't call tools anymore — so it can't save to Engram either. The escape hatch: grab the share link and run one CLI command. Every word rescued, searchable forever.
· 5 min read · community, obsidian, integrations, api
Engram Vault mirrors your AI memories into Obsidian as real markdown notes — synced, tagged, graphable, searchable. Built by a customer on our public API, no permission needed. And his one feature request shipped: the update API is live.
· 6 min read · vault, security, secrets, cli, how-it-works
Engram's secrets vault is zero-knowledge: keys are generated on your machine, values are AES-256-GCM encrypted before upload, and the server only ever stores ciphertext. How it works, and exactly which surfaces get it.
· 6 min read · claude-code, tokens, context-window, mcp, how-to
A long Claude Code session re-reads 100k+ stale tokens on every turn. Retrieving the same context from Engram costs about 2,000 — we measured it. Here's the math and the save → /clear → recall workflow.
· 5 min read · import, chatgpt, how-it-works, deduplication
Afraid re-importing your ChatGPT export will double everything? Engram deduplicates by conversation — re-upload as often as you like, and only genuinely new conversations get added. Here's exactly how it works, honestly.
· 8 min read · memory, codex, claude-code, verbatim, comparison
ChatGPT, Claude, and Codex all ship "memory" now — but every one of them stores a compressed summary, not your actual words. Here's the evidence from Codex's own database, why they all work this way, and what verbatim memory looks like instead.
· 7 min read · chatgpt, export, backup, tutorial
Your ChatGPT conversations are locked in one app. Here's how to export your full history, back it up permanently, and make every conversation searchable from ChatGPT, Claude, Cursor, and any AI you use.
· 7 min read · chatgpt, cross-platform, mcp, tutorial
Connect Engram to ChatGPT in 30 seconds. Then connect it to Claude Code, Cursor, or any MCP client. Research in one tool, execute in another — same memory, zero copy-paste.
· 8 min read · team, agents, shared-memory, architecture
A seat on Engram's Team plan doesn't have to be a person. It can be a deploy bot, a support agent, a code reviewer — any AI agent that needs shared context with the rest of your team.
· 7 min read · claude-code, mcp, tutorial
Claude Code resets its context every session. Here's how to add cross-session memory using Engram and the Model Context Protocol, so your agent remembers past decisions, investigations, and context.
· 8 min read · comparison, memory, mcp
A fair comparison of three agent memory servers — Mem0, Zep, and Engram — covering storage models, protocols, search, pricing, and when to pick each.
· 6 min read · cursor, mcp, tutorial
Cursor resets context every session. With one MCP config block, you can give it memory that persists across sessions, projects, and machines — using Engram.
· 9 min read · engineering, cloudflare, architecture
Why we picked D1 + Vectorize + Workers AI over the obvious stacks for Engram, what the latency math looks like, and where the edges of each service actually are.