Mnemo

A local-first MCP server that gives AI agents persistent memory and retrieval over large document corpora. Instead of dumping entire files into the context window, mnemo pre-digests them into searchable chunks and serves only what's relevant.

Named for Mnemosyne, the Greek Titaness of memory.

How It Works

Mnemo combines vector similarity search (sqlite-vec) with keyword search (FTS5) using Reciprocal Rank Fusion to find the most relevant chunks from your corpus. Embeddings are generated by a remote Ollama instance.

Agent ──MCP stdio──▶ mnemo ──▶ Ollama (embeddings)
                        │
                        ▼
                     SQLite (sqlite-vec + FTS5)

Two modes of storage, separated by "spaces":

• product — pre-ingested document corpus (markdown, JSON, plain text)
• default — episodic memory accumulated by the agent over time

MCP Tools

Tool	Description
recall	Hybrid search across memory and corpus
remember	Store a new memory (auto-classified by type)
forget	Soft-delete a memory by chunk ID
memory_status	Health check and per-space chunk counts

Quick Start

Prerequisites

• Go 1.22+
• An Ollama instance with an embedding model pulled (e.g., nomic-embed-text)

Build

make build

Ingest a corpus

mnemo ingest --ollama-host https://your-ollama-host ~/path/to/docs/

Register with Claude Code

claude mcp add --scope user \
  -e MNEMO_DB_PATH=/home/you/.mnemo/mnemo.db \
  -e MNEMO_OLLAMA_HOST=https://your-ollama-host \
  -- mnemo /path/to/mnemo serve

Agent instructions

Add to your CLAUDE.md:

Persistent memory and product knowledge are available via recall, remember, forget. Call recall with spaces:["product"] to query the product corpus. Call recall (all spaces) at task start to restore prior decisions. When a decision is made or something is learned, remember it as one concise standalone statement. Use forget only when information is explicitly superseded.

Configuration

All configuration via environment variables (prefix MNEMO_) or flags:

Variable	Default	Description
MNEMO_DB_PATH	./mnemo.db	Path to SQLite database
MNEMO_OLLAMA_HOST	https://ollama.cluster.collins.is	Ollama API URL
MNEMO_EMBED_MODEL	nomic-embed-text	Embedding model name
MNEMO_EMBED_DIM	768	Embedding dimension (must match model)
MNEMO_OLLAMA_TIMEOUT	30s	Request timeout for Ollama
MNEMO_CHUNK_TARGET_TOKENS	400	Target chunk size in tokens
MNEMO_LOG_LEVEL	info	Log level (debug, info, warn, error)

Tech Stack

• Go — pure-Go build, no CGo
• modernc.org/sqlite with sqlite-vec — vector search
• SQLite FTS5 — keyword search with BM25 ranking
• MCP Go SDK — stdio transport
• Ollama — embedding generation via /api/embed

Acknowledgments

Mnemo's hybrid search algorithm, schema model, and classification heuristics are adapted from Memory Vault by mihaibuilds. The idea to build a local RAG memory server came from Yadullah Abidi's article "I fixed Claude's memory problem with a Postgres database..." (MakeUseOf, 2026-06-16).

License

MIT