Recoleta
GitHub
Source note

I reverse-engineered Claude Desktop's storage to give it memory

Authors
foufouadi
Source
Hacker News
Published
Collection
Software Intelligence
Code IntelligenceMemory Augmented LLMLocal First AIMCPRetrieval Augmented Generation
Open GitHubMarkdown source

Summary

Mnemos gives Claude Desktop a local memory layer by reading Claude's on-disk conversation storage, indexing it, and exposing search through MCP. It aims to avoid long-context slowdown and confusion by retrieving relevant past fragments instead of keeping everything in one active chat.

Problem

  • Claude Desktop has no history or memory API, so past conversations are hard to reuse in a structured way.
  • Very long chats can degrade model behavior as context grows, with slower responses, more confusion, and more hallucinations; the project describes this as context rot.
  • Users who want persistent memory with privacy may not want cloud sync or external API calls.

Approach

  • Mnemos reverse-engineers Claude Desktop's local Chromium storage, watches files in real time, decompresses stored data, and parses conversation records from active sessions and history.
  • It embeds messages locally with the ONNX version of MiniLM-L6-v2, then stores text and vectors in SQLite using FTS5 for keyword search and vector blobs for semantic search.
  • It exposes retrieval back to Claude through an MCP server over JSON-RPC, so Claude can query relevant past snippets on demand.
  • Search uses hybrid retrieval: BM25 keyword ranking from SQLite FTS5 plus cosine-similarity semantic ranking, merged with Reciprocal Rank Fusion.
  • v1.1 adds a local GUI that lets users browse, search, and visualize conversation history with UMAP and K-Means in a 3D view.

Results

  • The system claims fully offline operation: no API calls, no cloud sync, and no data leaving the machine.
  • It claims real-time indexing with OS file watchers and semaphore debouncing, with zero polling overhead.
  • Embeddings use MiniLM-L6-v2 with 384-dimensional vectors, and the visualization projects those embeddings with UMAP plus K-Means clustering.
  • The paper excerpt gives no benchmark numbers for retrieval quality, latency, memory savings, hallucination reduction, or user outcomes.
  • The strongest concrete claim is architectural: hybrid search with Reciprocal Rank Fusion gives better recall than semantic search alone or keyword search alone, but no measured comparison is included in the excerpt.

Link