11.03.2026 17:07
👍 0
🔁 0
💬 0
📌 0
I'm writing about this architecture — context graphs → promise graphs → social memory — in my Age AI: Pocket Knowledge Graphs book series.
📚 Vol 1 is out. Next volumes go deep into what lies beyond context graphs.
#AgenticAI #MultiAgentSystems #KnowledgeGraphs #PromiseTheory #AIMemory #SovereignAI
Without social memory: a random collection of tools.
With social memory: something closer to an organization.
Earned trust. Learned patterns. Accumulated relational knowledge.
This isn't humanizing agents. It's building systems that actually collaborate.
The human question:
Agents interact with humans too — and the interface is fundamentally different.
Should agents maintain "working relationships" with humans? Learn their style, risk tolerance, patterns?
Not emotional intelligence. Operational intelligence.
The holacratic dimension:
Can we delegate to this agent? Can we form temporary teams? Assign roles? Dissolve when done?
Social memory provides the relational knowledge to make these decisions.
What social memory tracks:
→ Trust boundaries — your agents must NOT whisper secrets to competitors
→ Capability profiles — observed performance, not advertised skills
→ Collaboration patterns — who works well together
→ Outcome verification — did they deliver what they promised?
Layer 3: Social Memory
The layer nobody is building yet.
If promise graphs tell you what happened BETWEEN agents, social memory tells you what to EXPECT next time.
Promise graphs reveal what action logs can't:
→ Intention-execution gaps
→ Dependency chains between agents
→ Conditional reasoning
→ Temporal arcs from commitment to resolution
Layer 2: Promise Graphs
Building on Mark Burgess's Promise Theory.
Not just what agents DID — what they COMMITTED to do.
The gap between promise and delivery is where real learning happens.
Layer 1: Context Graphs
What happened? Decisions made, data consumed, actions taken.
The audit log. Retrospective. Factual.
Necessary but nowhere near sufficient.
The hard part: agents now talk to OTHER agents.
Not just human ↔ agent anymore. It's agent ↔ agent ↔ agent.
And that demands completely different memory architecture.
