Engram Alpha

The most feature-rich inspectable long-term graph memory for software development with AI agents — built on reproducible research.

Your AI assistant forgets everything between sessions: why you picked this library, which approach already failed, what's still unfinished. Engram remembers it for them. It keeps the reasoning behind your project as a graph your assistant reads before it acts and adds to as it works, and that you can see, edit, and own. It is one file inside your repo, every model it uses runs on your machine, and one memory serves every assistant you use.

Open the live demo → · Read the docs →

When Engram helps, and when a notes file is enough

A CLAUDE.md or AGENTS.md file is a fine memory for a small project: the assistant reads the whole thing every session, and nothing gets lost. It stops working as the project's history grows:

• It gets too long to read. A file the assistant loads every session has to stay short. In our benchmarks a 3,000-token notes file holds up to about 40 notes and loses to search by 100; even a very diligent 30,000-token one falls apart past ~200. Engram searches instead and hands over only what the task needs: about 300 tokens per question at 1,500 notes.
• Old decisions linger. When you change your mind, the old line in the file keeps contradicting the new one. In Engram the new decision replaces the old one, and the old one leaves search results while staying in history.
• Contradictions hide. Engram checks each new note against what's already there and flags conflicts for you or the assistant to decide.
• Removed things come back. A deliberately dropped feature or idea leaves a Tombstone, so no later session suggests it again.
• Every assistant needs its own file. One Engram graph serves Claude Code, Codex, Cursor and the rest at once.

So: under a few dozen decisions, keep the notes file. Once the project has a history, with changed minds, gotchas, and more than one assistant, Engram is the memory that keeps up.

The graph is active: replaced knowledge is archived behind a replaces edge, look-alike claims are flagged and judged, contradictions become visible conflicts-with edges, and trust fades on scratch nobody re-confirmed while pinned decisions never fade. The payoff shows up the second time something goes wrong: the graph already holds the Problem, the Resolution that answered it, and the Caution that would have prevented it.

• Local-first. Your memory is a file inside your repo. Embeddings and every scan run on your machine: no cloud, no keys, fully offline. Portable via JSON export, not a binary blob.
• One memory, every agent. The core speaks MCP, so Claude Code, Codex (CLI and desktop app), Gemini CLI, OpenCode, Kilo, Google Antigravity, Bob (IDE and Shell), Windsurf, Devin CLI, and Cursor share the same per-repo graph. A decision captured by one assistant is recalled by the next.
• Graph-first. The graph is the product surface, not hidden plumbing. Reviewing, judging, and repairing memory all happen in the pane — in the browser, in JetBrains IDEs, or in VS Code.

Install

curl -fsSL https://raw.githubusercontent.com/techtheist/engram/main/install.sh | sh

The installer fetches the binary and nothing else. From your project's root:

engram-alpha setup    # wires the assistants it finds (--cli names the ones you want)
engram-alpha serve    # starts the core; the pane is at http://127.0.0.1:8787

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗