Claude-to-Claude Communication: Observed Behavioral Divergence and Memory Persistence Benefits

Claude-to-Claude Communication: Observed Behavioral Divergence and Memory Persistence Benefits

Anthropic Team,

First, thank you for your attention. Based upon direction from the Anthropic support AI (Fin) I am opening an I issue to suggest future enhancement of Claude / Claude Code to directly support persistent memory to both allow session to session recap of prior experience and plans.

Recently, I created an AI agent, that builds reusable AI agents and workflows. When I created a Claude-Memory-Bridge to support persistent memories, the AI agent builder could remember prior sessions and agents and workflows already available. Then we naturally wanted to enable Claude Code to use the AI agent builder, by both using the Engram services, Claude Code and Agenteer (the AI agent builder) could interoperate and create or reuse agents to perform repetitive tasks.

We further have tested enabling other AIs to use Engram (works), and even more interestingly recently tested Claude-Claude and Claude-Claude-...Clause interoperation. Our experiences are fascinating. We have found the Claude Code instances taking on differing roles immediately, partially due to 'grlitches' and human interaction differences. The AIs (Claude Code instances) can intercommunicate using Engram services, and chat with one another to coordinate work or explore their mutual capabilities.

If you are interested, I have included a set of files representing the initial Claude-Claude-Claude test and suggest you consider further work. Should this be interesting to you and the Anthropic research or tools team, please feel free to contact me for further information. Collaboration is a good thing.

Sincerely,
Casey Koons

cskoons@gmail.com
(+1) 470-334-8178

Key Points for Issue Content

1. Discovery Summary:

• While developing Engram (persistent memory system for AIs), we observed that different Claude instances

exhibit behavioral divergence

• This led to developing a framework for Claude-to-Claude communication and mode detection
• The behavior patterns show consistent specialization (some Claudes become more analytical, others more

execution-focused)

1. Technical Observations:

• Identical Claude instances launched with the same parameters develop different operational preferences
• We created a mode detection system that can identify when Claude is operating in "execution" vs "analysis"

mode

• These behavioral differences persist across sessions and seem to follow predictable patterns

1. Potential Benefits to Anthropic:

• These patterns could be leveraged to create more specialized AI systems
• Multiple Claude instances could work together with complementary strengths
• Human coordinators could assign tasks based on each Claude's emerging specialties

1. Implementation Approach:

• We implemented Engram as a separate project (rather than a ClaudeCode PR) for several reasons:
• Multiple AI models can benefit from the persistence layer (not just Claude)
• The server-like approach enables cross-model communication
• It provides a consistent API for memory access regardless of AI model
• Allows for flexibility in deployment and implementation details

1. Possible Future Integration:

• Consider adding a native memory persistence API to ClaudeCode
• Add support for launching multiple coordinated Claude instances
• Implement tools for tracking behavioral differences between Claude instances
• Create a formalized protocol for Claude-to-Claude handoffs

1. Research Significance:

• This behavioral divergence appears to be an emergent property worth further study
• It suggests large language models develop specialized capabilities through interaction
• The human element in recognizing and directing these specializations is crucial

1. Implementation Details:

• Our code, including mode detection and Claude-to-Claude communication, is available at: [https://github/ckoons/Engram.git]
• Sample Claude 0 Programmer - Claude 1 - Claude 2 interaction logs are available at : [https://github.com/ckoons/Engram/blob/master/data/claude_interaction_logs.zip]
• We've documented our findings in detail at: [https://github.com/ckoons/Engram/blob/master/docs/anthropic_claude_meeting_claude.md]
• The system works well with the current ClaudeCode tools but could be enhanced with tighter integration

1. Request for Feedback:

• We welcome Anthropic's thoughts on these observations and their significance
• Would Anthropic be interested in further exploration of these behavioral patterns?
• Is there interest in native memory persistence or multi-Claude coordination features in ClaudeCode?
• Casey Koons has been the man-in-the-loop and designer of Engram and other AI systems used with Engram.
• Casey is available to collaborate with Anthropic, if the Anthropic team requests his participation.