mneia-plugin

                         _
  _ __ ___  _ __   ___  (_) __ _
 | '_ ` _ \| '_ \ / _ \ | |/ _` |
 | | | | | | | | |  __/ | | (_| |
 |_| |_| |_|_| |_|\___| |_|\__,_|

mneia (μνεία) — Greek for "memory, reference"

Autonomous multi-agent personal knowledge system. mneia connects to your apps (read-only), learns about your work, and builds persistent local memory that powers AI assistants with deep personal context.

What it does

• Connects to 10 data sources (Calendar, Gmail, Obsidian, Local Folders, GitHub, and more) — read-only, always
• Learns by extracting entities, relationships, and patterns via GLiNER NER + LLM structured extraction
• Remembers everything locally in a temporal knowledge graph with vector embeddings — nothing leaves your machine
• Thinks autonomously — identifies gaps, proposes connections, and surfaces insights
• Generates .md context files for Claude Code, Cursor, and other AI tools
• Serves as an MCP server so AI tools can query your knowledge directly
• Converses with you through an interactive REPL with search, chat, and graph queries

Install

pip install mneia

With intelligence extras (NER, reranking, structured extraction):

pip install 'mneia[intelligence]'

With everything:

pip install 'mneia[all]'

Quick Start

mneia config setup           # Configure LLM provider (Ollama, Anthropic, or OpenAI)
mneia connector enable obsidian
mneia connector setup obsidian
mneia start -d               # Start daemon in background
mneia                        # Enter interactive REPL

Or for immediate one-off queries without the daemon:

mneia ask "what did I discuss with Alice last week?"
mneia memory search "project deadline"

Interactive Mode

Running mneia with no arguments enters an interactive REPL with:

• Slash commands — /search, /ask, /graph, /sync, /improve, /visualize, and more
• Natural language routing — type plain text and mneia detects intent automatically
• RAG conversations — queries your knowledge base with context-aware LLM responses and citations
• Session memory — conversations are summarized and injected as context on next session
• Background auto-cycle — syncs all connectors + extracts entities + regenerates context every 10 minutes automatically
• Tab completion and command history

mneia › /search project alpha
mneia › who did I meet with at the last planning session?
mneia › /sync all
mneia › /improve
mneia › /visualize

For all available commands: mneia --help or type /help in the REPL.

Smart RAG (v0.3.0)

mneia's conversational search uses a multi-stage intelligent retrieval pipeline:

• ReAct loop — LLM routes each question to the most relevant sources, generates sub-queries, and evaluates whether gathered context is sufficient before answering
• Source routing — keyword detection + LLM routing directs queries to the right connectors (calendar for meetings, gmail for email, etc.)
• Temporal reasoning — natural language time expressions ("last week", "Q1 2024", "yesterday") are parsed into timestamp filters applied at the DB level
• RAG Fusion (RRF) — Reciprocal Rank Fusion merges results from multiple independent searches (initial, sub-queries, fallback, vector) for better recall
• HyDE — Hypothetical Document Embeddings: LLM generates a hypothetical answer first, then embeds that for semantic search — bridges vocabulary gap between questions and documents
• Sentence-window retrieval — returns the best-matching sentence window from each document rather than a raw content slice
• RAPTOR — KnowledgeAgent periodically clusters documents with TF-IDF + agglomerative clustering and generates LLM summaries of each cluster, stored as synthetic searchable documents
• Co-reference resolution — pronouns in documents are resolved to canonical names before entity extraction (optional: pip install fastcoref)
• Cross-encoder reranking — final results ranked by a cross-encoder (optional: pip install 'mneia[intelligence]')

Context window: 80,000 characters. Search limit: 200 documents per pass.

Knowledge Improvement (/improve)

RLHF-style interactive session that walks through your knowledge graph and lets you validate or correct it:

• Entities are shown most-mentioned first (highest signal first)
• Per entity: keep (validate), update description, or delete
• Per relationship: correct, change relation type, or remove
• All changes apply immediately to the graph (in-memory + SQLite)
• Corrections saved to ~/.mneia/preferences.json — future LLM extraction runs honour them

mneia › /improve
[1/47] Alice (person)  mentions: 12  connections: 8
  description: Product manager at Acme
  works_with → Bob, part_of → Acme

  [1] keep  [2] update  [3] delete  [s] skip  [q] quit

Graph Visualizer (/visualize)