# Cursor-memory — Persistent, Searchable Memory for Cursor AI

**URL:** <https://forum.cursor.com/t/cursor-memory-persistent-searchable-memory-for-cursor-ai/156344>\
**Category:** Built for Cursor\
**Tags:** mcp, context\
**Created:** [March 31, 2026, 3:09pm UTC](https://forum.cursor.com/t/cursor-memory-persistent-searchable-memory-for-cursor-ai/156344 "2026-03-31T15:09:27Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![tranhuucanh](https://sea3.discourse-cdn.com/cursor1/user_avatar/forum.cursor.com/tranhuucanh/32/103677_2.png) [@tranhuucanh](https://forum.cursor.com/u/tranhuucanh)\
**Post date:** [March 31, 2026, 3:09pm UTC](https://forum.cursor.com/t/cursor-memory-persistent-searchable-memory-for-cursor-ai/156344/1 "2026-03-31T15:09:27Z")

</div>

Hi everyone,

I built an MCP server that gives Cursor persistent memory across chat sessions.

## The problem

Every new chat starts fresh. Decisions, architecture context, debugging sessions — all gone. You end up re-explaining the same things over and over, or maintaining .md files manually.

## What cursor-memory does

It lets you save and recall memories using simple commands in chat:

- `/memo` — AI summarizes the conversation into a structured memo (Decisions → Key Details → Context → Next Steps), auto-tags it, and stores it locally
- `/recall [query]` — search your memories by meaning
- `/forget [query]` — delete memories you no longer need
- **Auto-recall** — AI detects when your question refers to something you’ve discussed before and searches automatically. No command needed.

## Key features

- **You control what’s saved** — nothing is saved without you saying /memo
- **Semantic search** — hybrid FTS5 + vector search. Finds by meaning, not just exact keywords
- **Repo-scoped + global** — repo memories stay isolated per project, global memories are available everywhere
- **Multilingual** — save in English, search in Japanese. 100+ languages supported
- **Handles long content** — long discussions are chunked into overlapping segments so nothing gets lost
- **100% local** — SQLite + local embeddings (multilingual-e5). No cloud, no API keys, no telemetry
- **Cross-platform** — macOS, Linux, Windows

## Installation

```bash
npm install -g cursor-memory

cursor-memory setup

```

That’s it. Two commands. It configures the MCP server, installs the AI behavior rules, and downloads the embedding model. Restart Cursor and you’re ready.

## How it works under the hood

- Storage: SQLite with FTS5 for full-text search
- Embeddings: multilingual-e5 models via ONNX Runtime (runs locally in a worker thread)
- Search: hybrid scoring — 40% keyword match + 60% vector similarity
- Long content is split into overlapping chunks (350 words, 50 word overlap) to stay within the model’s 512-token window
- Three model sizes available: small (~50MB), medium (~115MB), large (~270MB)

## Links

- GitHub: [GitHub - tranhuucanh/cursor-memory: Persistent, searchable memory for Cursor AI — You control what your AI remembers. · GitHub](https://github.com/tranhuucanh/cursor-memory)

- npm: [https://www.npmjs.com/package/cursor-memory](https://www.npmjs.com/package/cursor-memory)

Would love feedback from the community. Especially interested in:

- Any issues on Windows/Linux (I primarily develop on macOS)

Thanks!