ALBS AI Memory Architecture

Outcome

95% context retention across session boundaries — measured over 4 months of production use

• 3-tier architecture: hot (sub-15KB), warm (996+ Obsidian files), cold (session transcripts)
• Automatic promotion/demotion based on access frequency (3+ refs in 7 days = hot, 30 days unreferenced = warm)
• 14 agents across 4 hosts sharing unified memory layer
• Semantic search across all tiers via mdvs (vector + BM25 + RRF)

The Problem

AI agents lose context between sessions. Critical decisions, client history, and operational patterns vanish. Multi-agent teams need shared memory that survives restarts, host migrations, and agent upgrades — without unbounded context windows.

The Architecture

Designed a 3-tier memory system: hot tier (MEMORY.md, sub-15KB, identity + active priorities), warm tier (Obsidian vault + briefs, retrieved on demand), cold tier (raw daily notes + session transcripts, explicit retrieval only). Automatic demotion: 30 days unreferenced moves to warm, daily notes older than 14 days archive. Every agent reads hot tier on startup and writes decisions to disk immediately.

The Solution

14 agents across 4 hosts (Hetzner, Northstar, Office Paled, Oracle) share a unified memory layer. Context retention measured at 95% across session boundaries. The Obsidian vault serves as the warm tier with 996+ interconnected markdown files. Memory search provides semantic retrieval across all tiers.

Stack & Role