Engram

Every AI coding assistant has the same annoying flaw: it wakes up with amnesia. Close the session, and whatever it learned about your codebase, your architecture decisions, your reasons for picking one library over another — gone. Engram, a new open-source project, tries to fix that by giving agents something closer to a brain, borrowing its name from the neuroscience term for a physical memory trace.

Under the hood it's refreshingly unglamorous: a single Go binary backed by SQLite with FTS5 full-text search. No Node.js, no Python, no Docker required. You install it with a one-line brew command, and it exposes itself through a CLI, an HTTP API, an MCP server, and a terminal UI built with vim-style navigation. Because it speaks MCP, it plugs into pretty much anything that supports the protocol — Claude Code, OpenCode, Gemini CLI, Codex, VS Code's Copilot, Antigravity, Cursor, Windsurf, and others. Setup is mostly automated: run a setup command for your agent, restart it, and the MCP wiring is done for you.

The memory system itself runs on 20 MCP tools split across categories: saving and updating memories, searching and retrieving them, tracking session lifecycles, surfacing conflicts, and general utilities like project stats and a doctor command for diagnostics. The basic loop is simple — an agent finishes meaningful work, calls mem_save with a title and some structured notes on what, why, and where, and Engram writes it into SQLite with full-text indexing. Next session, the agent searches that memory store instead of starting from zero.

For teams working across machines, Engram ships a git-based sync mechanism that moves memories as compressed chunks rather than raw files, avoiding merge conflicts. Cloud sync is available too, but it's explicitly opt-in and always scoped to a specific project — local SQLite stays the source of truth, and the cloud is just replication.

The more interesting stuff is still in beta: a conflict-surfacing system that flags when two saved memories contradict each other, plus a semantic scan mode that hands ambiguous cases to your existing Claude Code or OpenCode subscription to judge whether one decision actually supersedes another — at no extra cost if you're already paying for that agent. It's an early, isolated beta with its own Docker setup and cleanup command, but it points at where this is headed: not just remembering what an agent did, but noticing when it changes its mind.