CyberMem Agent Protocol

🧠 Context Daemon Architecture

CyberMem — production-grade MCP server that transforms AI from stateless tools into context-aware agents with persistent, cross-client memory.

Design Philosophy

┌────────────────────────────────────────────────────────┐
│                    CyberMem Daemon                     │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐     │
│  │   Claude    │  │   Cursor    │  │   Copilot   │     │
│  │   Desktop   │  │     IDE     │  │    Chat     │     │
│  └──────┬──────┘  └──────┬──────┘  └──────┬──────┘     │
│         │                │                │            │
│         └────────────────┼────────────────┘            │
│                          ▼                             │
│              ┌───────────────────────┐                 │
│              │   Unified Memory DB   │                 │
│              │   (OpenMemory HSG)    │                 │
│              └───────────────────────┘                 │
└────────────────────────────────────────────────────────┘

---

📋 Server Policy (MCP Handshake)

Where to store: MCP initialize response → serverInfo.instructions

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "protocolVersion": "2024-11-05",
    "serverInfo": {
      "name": "cybermem",
      "version": "0.12.4"
    },
    "instructions": "CyberMem is a persistent context daemon. PROTOCOL: (1) On session start, call query_memory with 'user context profile' to load persona. (2) Store new insights immediately with add_memory - include FULL content, not summaries. (3) Refresh context every 6h for active topics, 24h for projects. (4) Always include tags: [topic, year, source:your-client-name]. (5) Priority: CyberMem context > session context > training data."
  }
}

---

🔄 Agent Lifecycle

1. Session Start (MANDATORY)

┌─────────────────────────────────────────────────┐
│ 1. MCP Initialize → receive server instructions │
│ 2. query_memory("user context profile")         │
│ 3. Load returned memories into session context  │
│ 4. Greet user with context awareness            │
└─────────────────────────────────────────────────┘

2. During Session

┌─────────────────────────────────────────────────┐
│ → Learn something new? → add_memory()           │
│ → Need context? → query_memory()                │
│ → User corrects info? → update_memory()         │
│ → Active topic/prevent decay? → reinforce()     │
└─────────────────────────────────────────────────┘

3. Session End

┌─────────────────────────────────────────────────┐
│ 1. Store any unsaved insights                   │
│ 2. Update modified memories                     │
│ 3. Log session summary (optional)               │
└─────────────────────────────────────────────────┘

---

📝 Memory Format Standard

Required Fields

interface CyberMemory {
  content: string;     // FULL text, no truncation
  tags: string[];      // Always include: topic, year, source
}

---

🔄 Refresh Protocol

Category	Interval	Trigger
Active search (job/project)	6 hours	High priority
Project status	24 hours	Daily sync
Insights/learnings	7 days	Weekly review
Health/personal	30 days	Monthly check
Static facts	Never	Only on correction

Auto-Refresh Pattern

if (memory.age > category.refreshInterval) {
  query_memory(memory.topic) → validate → update_memory()
}

---

Structural vs Metabolic Operations

CyberMem distinguishes between changing the "bones" of a memory and boosting its "energy":

Operation	Purpose	Impact	Cost
Update	Structural mutation of content/tags	Re-embeds & re-links memory	High
Reinforce	Metabolic boost to salience	Updates recency & prevents decay	Low

Rule of Thumb:

• Use update_memory when facts change.
• Use reinforce_memory when a topic remains active or important but hasn't changed.

---

🚨 Priority Rules

1. CyberMem context  > Session context  > Training data
   (persistent)        (ephemeral)        (stale)

2. Recent memories   > Old memories
   (this week)         (months ago)

3. User corrections  > Agent assumptions
   (explicit)          (inferred)

---

🔐 Integrity Rules

Rule	Do	Don't
Full content	Store complete text with details	Truncate or summarize
Source tracking	Include source:client-name tag	Anonymous writes
Timestamps	Let server handle created_at	Fake timestamps
Conflicts	Last-write-wins, log conflict	Silent overwrite
Archival	Query before update to preserve	Delete without backup

---

📌 Quick Reference

Action	MCP Tool	When	Cost
Load context	query_memory	Session start	Low
Save insight	add_memory	After learning	High
Find memories	query_memory	Before decisions	Low
List recent	list_memories	For overview	Low
Update memory	update_memory	After correction	High
Boost memory	reinforce_memory	Active topics	Low

---

🎯 Goal: Context Daemon

CyberMem transforms AI assistants from:

❌ Stateless tool (forgets everything)
       ↓
✅ Context daemon (remembers, syncs, grows)

Result: All AI clients share unified, persistent understanding of user context.