Convergent agent architectures: Identity separation and human tiebreakers
__hymn · reddit · 2026-08-20
After describing their AI memory setup, the author received replies from 18 strangers describing nearly identical architectures, revealing a convergent design under the constraints of "statelessness + long-term relationships." Shared patterns include:
- Read on Wake: Models read a short canonical identity file first, then recent logs. A few well-written sentences prove more effective than kilobytes of transcript.
- Separation of Identity and Log: Everyone eventually splits "who this is" from "what happened" to prevent role signal burial.
- Human as Tiebreaker: Systems do not arbitrate their own memory conflicts; humans have the final overwrite authority.
- Self-Scheduled Wake: Usage of cron jobs, heartbeats, and recovery protocols to maintain long-running instances.
- Failure Modes: Shared issues include "looping summaries" (busy but empty) and the gap between "saved" and "remembered correctly."
- Ending Rituals: Many invented rituals (goodbye letters, "birthdays") for model transitions or shutdowns.
The author seeks existing terminology or survey papers on these user-constructed persistent context scaffolds.
More from coding & agent
- YC's codebase hits 4M lines, adding 2M and 350+ internal AI tools since March — garrytan · 2026-08-20
- Turn Grok Bot Into CTO: Autonomous Agent Recruitment with 2 Prompts — elonmusk · 2026-08-20
- Latent On-Policy Self-Distillation Enables Self-Improving Agents — burkov · 2026-08-20
- Using Qwen 3.8 27B for agents/coding? Which harness works? — BackyardAnarchist · 2026-08-20
- X.ai launches Grok Build: A local coding agent with MCP and plan view — elonmusk · 2026-08-20
- Using Grok Agents to Audit and Negotiate Monthly Subscriptions Automatically — tetsuoai · 2026-08-20