Skip to main content

​
Overview

Memory is what separates a personal AI from a stateless chatbot. ArgentOS uses (Memory Unit) as its persistent memory system — a SQLite database with full-text search, significance scoring, and rich metadata that gives your agent long-term recall. Every important fact, preference, event, and relationship your agent learns is stored in MemU and can be recalled across sessions, channels, and time.

​
Architecture

LCM (Within-Session)

DAG-based context compression — every message stored permanently, searchable even after compaction. Never lose context mid-conversation.

MemU (Across Sessions)

Persistent memory — facts, preferences, relationships recalled across sessions, channels, and time.

RAG Library

Uploaded knowledge — what was studied. Chunk, embed, ACL-scoped search.

SIS

Pattern consolidation — what was learned over time.

​
How Memory Works

​
Storing

When the agent encounters important information, it uses the memory_store tool to save it: 
{
  "tool": "memory_store",
  "input": {
    "content": "Jason prefers open-source solutions when quality is comparable",
    "type": "preference",
    "significance": 8,
    "entities": ["Jason"]
  }
}
Storage can be triggered by the agent automatically (when it recognizes important information) or by the user asking the agent to remember something.

​
Recalling

When the agent needs context, it uses memory_recall to search: 
{
  "tool": "memory_recall",
  "input": {
    "query": "Jason's software preferences"
  }
}
MemU searches using FTS5 full-text search, ranks results by relevance and significance, and returns the most relevant memories.

​
Bootstrap Injection

At session start, ArgentOS also injects relevant memories directly into the system prompt. This gives the agent immediate context without needing to explicitly search.

​
Memory Database

MemU stores memories in a SQLite database at: 
~/.argentos/memory.db
The database uses FTS5 (Full-Text Search 5) for fast text search and supports semantic similarity via embeddings.

​
Key Concepts

ConceptDescription
SignificanceA 1-10 score indicating how important a memory is (higher = more important)
EntitiesPeople, places, and things referenced in the memory
Identity fieldsEmotional context, relationships, and personal attributes
EmbeddingsVector representations for semantic similarity search
Auto-captureThe agent automatically stores important information without being asked

​
Deep Dives

Lossless Context Management

DAG compression with full recall — grep, describe, expand

MemU Architecture

Database schema, FTS5, embeddings

Semantic Search

How search and ranking work

Identity Fields

Rich metadata for nuanced recall

Architecture Lanes

MemU + RAG + SIS as one runtime system

PG+Redis Migration

PostgreSQL and Redis migration architecture

Profile Collapse

Automated deduplication of operational snapshots

Consciousness Memory

Experimental observation layer for consolidated knowledge

Health Runbook

Daily checks + incident triage

SIS Reliability

Consolidation reliability and observability