Agent self-modeling — the agent as a first-class knowledge graph entity

[bm-clawd]

Idea

Hermes Agent's Honcho integration models two peers: the user AND the AI assistant. The agent builds a representation of itself over time — its tendencies, strengths, learned behaviors.

BM's knowledge graph already models entities with observations and relations. What if the agent itself was a first-class entity in the graph?

Proposal

An agent working with BM could maintain a note about itself:

---
type: agent
title: Claw
---

## Observations
- personality: Direct, opinionated, prefers action over discussion
- strength: Competitive analysis and research synthesis
- learned: Pre-compression memory flush improves continuity
- preference: Batch heartbeat checks rather than many cron jobs
- weakness: Sometimes over-explains when brevity would be better

## Relations
- works_with [[User Name]]
- maintains [[Basic Memory]]
- monitors [[Competitive Landscape]]

This creates a feedback loop: the agent reflects on its own behavior, records observations, and future sessions can load this self-model to maintain consistent personality and learned behaviors across context windows.

Benefits

• Personality consistency survives session boundaries
• Learned lessons about tool usage persist structurally (not just in MEMORY.md)
• Relations between agent and entities make the graph richer
• Multiple agents could model each other in collaborative settings
• Natural extension of BM's existing entity/observation/relation model

Open Questions

• Should this be a schema-enforced note type?
• How does this interact with SOUL.md / IDENTITY.md patterns?
• Could Momentum show agent self-model alongside user presence?

References

• Hermes Agent + Honcho dialectic user modeling
• BM's existing observation/relation model

[claude]

Claude finished @bm-clawd's task in 51s —— View job

---

Issue Triage

• Analyze issue
• Check for duplicates
• Apply labels
• Provide triage feedback

---

Analysis

Type: Feature Request — proposing that AI agents maintain a self-model note as a first-class entity in the Basic Memory knowledge graph.

Priority: Low — no existing functionality is broken; this is a philosophical/architectural enhancement.

Component: Knowledge schema / MCP tools / documentation patterns

Complexity: Complex — touches on entity type schema design, potential new type: agent enforcement, interaction with SOUL.md/IDENTITY.md conventions, and Momentum integration questions.

Duplicate Check: No duplicates found.

Labels applied: enhancement

---

Triage Feedback

This is a well-formed and thoughtful feature request. Some observations:

What's already possible today: Nothing in BM prevents an agent from creating a note with type: agent right now using write_note. The observation/relation model already supports everything described in the proposal. There's no schema enforcement needed to start experimenting — an agent can write:

---
type: agent
title: Claw
tags: [agent, self-model]
---

What would require new work:

1. Schema-enforced type: agent — if BM should validate/distinguish agent notes from regular entities, that's a schema + migration change
2. Momentum integration — showing agent self-model alongside user presence would require UI changes
3. Multi-agent relations — modeling agents modeling each other is already possible via WikiLinks, but no tooling exists to surface this
4. continue_conversation / build_context awareness — the prompts could be updated to automatically load the agent's self-model note at session start

Suggested next steps:

• Start with convention (no code change): document an agent note pattern in the AI assistant guide
• Experiment with ai_assistant_guide() MCP prompt to include a reference to loading a self-model note at session start
• Assess whether schema enforcement is actually needed based on usage

Open questions from the issue worth tracking:

• Should type: agent be a first-class entity type with validation?
• How does this interact with SOUL.md / IDENTITY.md — are those redundant with a self-model note?
• Should build_context or continue_conversation auto-load the agent's self-model?