SI Agent News
DevelopingClaudeOpenAIOpen modelsAgents

Memro, an early open-source project, gives coding assistants a shared memory

A new Show HN project aims to let Claude Code, Cursor and Codex remember your project between sessions. It is promising but very early, with strings attached.

500questions in the LongMemEval benchmark Memro reports using

The short version

Why it matters

Re-explaining your database, naming rules or past decisions at the start of every session is a real drain, and a memory layer that several tools can read reduces dependence on any single assistant. The catch is that you would be trusting a young project with notes about your work. Check where the memory lives, who can read it, and whether you can delete it before you rely on it.

What Memro is

Memro is a project that gives AI coding assistants a memory that outlasts a single chat session. As github.com reported through its Show HN listing, it targets three popular tools: Claude Code, Cursor and Codex. Show HN is a section of the Hacker News forum where makers present their own projects.

The pitch is simple. Today, most coding assistants start fresh each time you open a session, so you repeat yourself. Memro keeps a store of facts and decisions outside the assistant and hands them back when relevant.

Our hosts flagged this as a project announcement by its makers. We have not tested it, and the post had only 2 points when we saw it. That is a sign of little attention so far, not a verdict on quality.

How the memory works

The assistants connect to Memro through MCP, the Model Context Protocol. That is a standard way for AI tools to plug into outside services. Memro also offers a plain web API, a command-line tool, and kits for Python, JavaScript/TypeScript and Go.

The README's example shows the idea. You tell your assistant to remember that the primary database is PostgreSQL 16 with pgvector. You restart the session and ask which database and extensions you use, and the assistant should answer from memory.

Behind that sits a server written in Go, with a PostgreSQL database and an add-on called pgvector. The add-on lets the database search by meaning rather than exact words. Memro combines that with ordinary keyword search and merges the two result lists using Reciprocal Rank Fusion, a method that blends rankings from different searches.

More than a notepad

The project says it goes beyond simple storage. Memories come in four types:

It also describes tools for spotting contradictory facts across sessions and marking older ones as superseded, plus a way to trace how a belief changed over time. Each agent gets its own identity, built on a kind of cryptographic key called Ed25519, created locally.

For evidence, the makers point to a benchmark called LongMemEval, run across 500 questions on recall over many sessions and updated knowledge. They also describe a test suite covering stale facts, contradictions and distractor results. These are the makers' own reports, and we have not checked them.

Setup, license and privacy questions

To try it, the README tells you to run a connect command from the command line and log in through a device login. For more control, you can run the whole server and database on your own machine using Docker. That keeps the database in your hands, which matters for a tool that remembers your work.

One detail deserves attention: the self-hosting steps say to add an OpenAI API key for hybrid search. The source does not explain what is sent to OpenAI, so that is a question to settle before storing anything sensitive. The README also recommends setting a long access token to protect the server's metrics pages, and warns in its logs if you do not.

The license is the other catch. Memro uses the PolyForm Noncommercial License 1.0.0, which is source-available rather than fully open. Personal study, hobby projects, research and testing are free. Anything commercial needs a paid Memro Cloud subscription or an enterprise license. The README also says some install routes, such as the Python and npm packages and the version 0.1.0 container images, apply once those are published, so parts may not be ready yet.

Who should care

Developers who juggle several assistants and tire of repeating context are the obvious audience. A shared memory also makes it easier to switch tools without losing your accumulated notes, which fits a wider lesson: keep what matters in places you control.

Everyone else can skip it for now. If you do experiment, start with a throwaway project, store nothing secret, and confirm you can export or delete what it keeps. A prototype with 2 points is a reason to be curious, not a reason to depend on it.

What to watch

Try it

If you are a developer, read the README's self-hosting section and check what the OpenAI key is used for. Then test on a hobby project, not client work.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by github.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news