AI coding tools always forget? SynaBun writes memory into the SQLite on your computer

Every time you open a new coding session, Claude Code, Codex, or OpenCode has to be told again about the project background, what was changed last time, and which pitfalls not to step on again. The author Dan Kornas introduced a project called SynaBun in a tweet, and its goal is to end this repeated explanation.

SynaBun is an open-source, self-hosted AI workspace and persistent-memory MCP server. Memories are stored in a local SQLite file, and connected tools read and write them through MCP tools such as remember and recall. There are no accounts, no cloud service, and no API key required. The project is open-sourced on GitHub under the Apache 2.0 license.

The problems it solves are very concrete: preserve project decisions and fix records across sessions; search supports both keywords and semantics, with semantic retrieval relying on an optional local embedding model; a local web interface visualizes memories as a 3D map that can be browsed, searched, edited, and deleted; Claude Code, Codex, and OpenCode can run side by side and share the same memory; edits have version history, deletions go to the recycle bin, and backups can be verified and restored.

There is a point here that is easy to overlook. The value of a memory tool lies not only in how much it remembers, but also in whether it can be retrieved, modified, deleted, and restored. Remembering something wrong, remembering something outdated, or remembering sensitive information will all pollute the next session. SynaBun puts editing, the recycle bin, and backups into the interface, which counts as confronting this problem head-on. One reply under the tweet stands out: Shane to the B said his agent remembered every decision in the project except the one where he said “don't touch production.” Constraints cannot rely on memory alone.

How to use: Node.js 22.19 or newer is required. Install and start:

npm install -g synabun
npm --prefix "$(npm root -g)/synabun/mcp-server" run setup:embeddings
synabun

The second command downloads the local semantic search model. It still works without installing it, but only keyword search remains. synabun starts the local interface at http://localhost:3344. On first launch it opens a setup wizard to connect Claude Code, Codex CLI, Gemini CLI, and OpenCode; for Cursor, Windsurf, or other MCP clients, copy the configuration snippet following the wizard. After a global install, MCP clients can use synabun mcp over stdio.

When trying it out, you can just say: remember that releases go through the release branch, not main. Then open another session and ask: do you remember the deployment process? Then open http://localhost:3344 to see the memory map. Keep synabun running while you work; Claude Code accesses memory through it. If something seems off, run synabun doctor.

Local-first has a cost. Neural Interface has no login and listens on all network interfaces; do not expose port 3344 to the public internet. Manually copying memory.db may miss in-progress writes; for backups, use the app's scheduled backups and verified restore. AI tools will still send prompts and recalled memories to their own model providers; optional features such as AI judgment, WhatsApp, Discord, and the browser will connect to their respective services. Desktop operation is currently macOS-only.

Project address: GitHub, website: synabun.ai. If you are interrupted by session resets every day, you can try local memory first; if you don't want to maintain a local service, backups, and port security, a cloud memory tool will be less hassle. This trade-off is worth thinking through before “whether AI can remember.”