Rememori

Rememori is a memory engine for AI agents written in pure TypeScript. It lets agents remember, recall, and forget information across sessions without requiring external infrastructure. The library is tiny, about 8KB gzipped, with zero dependencies, so it slots into existing projects without bloating your bundle or complicating your deploy.

The API has three primary operations. Remember stores text along with embeddings, tags, and extracted entities. Recall retrieves relevant memories by combining cosine similarity, knowledge graph connections, user feedback, importance scores, and time decay. Forget removes memories you no longer need. The recall system is where most of the intelligence lives, because it ranks memories across multiple signals rather than just doing a naive vector search.

Entity extraction happens automatically. When you store a memory, Rememori identifies entities and builds a bipartite knowledge graph. That means recall can find relevant memories not just by semantic similarity but by shared entities. If two memories mention the same person or project, they get a boost even if the text itself isn't similar. This is useful for agents that need to connect information across conversations.

Storage is pluggable. In Node or Bun, it defaults to file-based storage. In the browser, it uses IndexedDB. For edge workers, there's a KV adapter. You can also bring your own storage if you need something custom. Embeddings are similarly pluggable: Ollama for local and private setups, any OpenAI-compatible endpoint, or a custom function you provide.

The library also works as an MCP server. You can connect it to Claude Code, Cursor, or Windsurf so your coding assistant gets persistent memory without any additional setup. That's one of the more immediate use cases: wire it up to your editor, and the assistant remembers context across sessions.

Version 0.7 added reinforce and demote functions. These let the agent self-correct its memory based on verifiable outcomes. If a recalled memory led to a good result, you reinforce it. If it led to a mistake, you demote it. Over time, the ranking improves without manual curation. The project is open source under MIT, available on npm, and runs anywhere JavaScript runs.