MemU Architecture

Overview

(Memory Unit) is ArgentOS’s built-in persistent memory system. It supports two backends:

PostgreSQL 17 + pgvector (recommended) — concurrent access, shared knowledge, HNSW vector indexes, tsvector full-text search with GIN indexes
SQLite (fallback) — single-process, local-only, FTS5 full-text search, stored at ~/.argentos/memory.db

The installer sets up PostgreSQL by default. SQLite is available as a zero-dependency fallback for simpler setups.

Database Schema

The core memory table contains:

Column	Type	Description
`id`	INTEGER	Primary key
`content`	TEXT	The memory content
`type`	TEXT	Category (fact, preference, event, relationship, etc.)
`significance`	INTEGER	Importance score (1-10)
`entities`	TEXT	JSON array of referenced entities
`emotional_context`	TEXT	Emotional tone or context
`source`	TEXT	Where this memory came from
`created_at`	TEXT	Timestamp of creation
`updated_at`	TEXT	Last update timestamp
`embedding`	BLOB	Vector embedding for semantic search

Full-Text Search

MemU uses different full-text search engines depending on the backend: PostgreSQL (default): tsvector with GIN indexes — supports ranked search, stemming, language-aware tokenization, and phrase matching. SQLite (fallback): FTS5 extension — boolean queries, phrase matching, prefix matching. Both backends support:

Boolean queries: preference AND open-source
Phrase matching: "infrastructure specs"
Prefix matching: config*
OR queries: telegram OR discord

Search Tips

Single keywords work best for broad recall
Use OR for multi-word queries to improve recall
Strip common stop words for better results
PostgreSQL handles stemming and language-aware tokenization automatically

Significance Scoring

Every memory has a significance score from 1-10:

Score	Level	Example
1-2	Low	Passing observations
3-4	Medium	Useful context
5-6	High	Important facts and preferences
7-8	Very High	Key personal information, critical decisions
9-10	Critical	Core identity, security-sensitive data

Higher significance memories are prioritized in search results and bootstrap injection.

Auto-Capture

The agent is trained to automatically store important information without being explicitly asked. When it detects a new fact, preference, event, or relationship in conversation, it calls memory_store proactively.

Embeddings

MemU supports vector embeddings for semantic similarity search. Embeddings are generated when memories are stored and used alongside full-text search for hybrid ranking. PostgreSQL: pgvector with HNSW indexes for fast approximate nearest neighbor search. SQLite: sqlite-vec extension for local vector search.

Final Score = text_relevance * weight + embedding_similarity * weight + significance_boost

Storage

PostgreSQL (default):

postgres://localhost:5433/argentos    # Database with pgvector extension
Redis on port 6380                    # Agent state, presence, pub/sub

SQLite (fallback):

~/.argentos/memory.db      # Main database
~/.argentos/memory.db-wal  # Write-ahead log
~/.argentos/memory.db-shm  # Shared memory

The database is accessed via getMemuStore() in src/memory/memu-store.ts.

Getting Started

Installation

Channels

Agents

Connectors

Tools

Concepts

Memory

Model Router

Overview

Database Schema

Full-Text Search

Search Tips

Significance Scoring

Auto-Capture

Embeddings

Storage

​Overview

​Database Schema

​Full-Text Search

​Search Tips

​Significance Scoring

​Auto-Capture

​Embeddings

​Storage

Overview

Database Schema

Full-Text Search

Search Tips

Significance Scoring

Auto-Capture

Embeddings

Storage