AI Agent Memory Architecture
How Our AI Agents Remember
A three-pillar system for giving AI agents persistent, searchable, self-maintaining memory — built on PARA, semantic search, and automatic decay.
The Problem
AI agents wake up with amnesia every session. They can't remember yesterday's conversation, your preferences, or what they were working on. Without a memory system, every interaction starts from scratch — and your agents never truly learn.
The Three Pillars
1
PARA
Structured Knowledge
Organize everything agents know into a clean hierarchy: shared facts, personal context, work projects, and family info. Facts are atomic and never deleted — only superseded.
2
QMD
Semantic Search
Every memory file is indexed for semantic search. Agents can find relevant context by meaning, not just keywords. Works as a fallback when primary memory tools are unavailable.
3
Decay
Self-Maintenance
A weekly automated process classifies memories as Hot, Warm, or Cold based on recency and access frequency. Cold memories get archived. The knowledge base stays lean.
Pillar 1 — PARA Knowledge Graph
🤝
S
Shared
- Cross-agent conventions
- People & contacts
- Calendar events
- Fact schema:
facts.json
🧑
P
Personal
- Health & fitness data
- Financial notes
- Tax documents
- Personal preferences
💼
W
Work
- Active projects
- Companies & contacts
- Business opportunities
- Client context
👨👩👧👦
F
Family
- Family members
- Important dates
- Shared activities
- Private context
{
"id": "vika-002",
"fact": "Vika and Zain work out together",
"category": "context",
"status": "active",
"relatedEntities": ["personal/health"],
"lastAccessed": "2026-02-01",
"accessCount": 3
}
Pillar 3 — Memory Decay
Hot
Active Memory
Recently accessed, highly relevant. Stays in working files. Agents see this every session.
Accessed within 7 days
Warm
Background Context
Still useful but not urgent. Available via search. Summarized during weekly grooming.
Accessed 8–30 days ago
Cold
Archived
Old context moved to archive. Retrievable via semantic search if needed, but out of the working set.
Not accessed in 30+ days
Pillar 2 — QMD Semantic Search
Query by Meaning, Not Keywords
Every markdown file is indexed. When an agent needs context, it searches by semantic similarity — "what was that conversation about the gym?" finds relevant facts even if the word "gym" isn't in the file.
Agent Query
"gym schedule"
→
→
Search Index
cosine similarity
→
Top Results
ranked snippets
The Agents
🔩
RivetPersonal Organizer
📿
Alpha TrionIslamic Knowledge
Each agent has its own memory space, but shares the knowledge graph. Privacy boundaries are enforced — some agents are air-gapped.
The Daily Rhythm
- Agent wakes up — reads its identity, the user profile, today's and yesterday's daily notes, and long-term memory file.
- Needs context? — searches the QMD semantic index across all its memory files and the shared knowledge graph.
- Learns something new? — writes it to today's daily note as a structured fact with categories and related entities.
- Weekly grooming (Sundays) — the decay script runs: Hot facts stay, Warm facts get summarized, Cold facts are archived. Daily notes older than 7 days move to the archive.
- Knowledge extraction — important facts from daily notes get promoted to the permanent knowledge graph, organized by PARA categories.