T H E W A L K

Design Technologist,
AI Infrastructure,
Prototyping.

← Back to main

Cross-Project Learning: Making the AI Persona Evolve

The Recap

In Milestone 1, I decoupled the AI persona from the project repository — a portable harness that can be mounted anywhere.

In Milestone 2, I moved project orchestration into a centralized registry, so the system knows which harness to boot based on the directory path.

But there was still a problem: persona stasis.

The Factory Reset Problem

Even with a portable identity, the AI was essentially starting fresh every time it moved to a new project.

The system correctly loads project-specific logs (what happened in the code), but the persona itself remained static. If the agent learned a new preference while working on Project A — say, how I like deployment flows structured — that knowledge was trapped in Project A’s logs. Moving to Project B meant a reset.

Milestone 3: Persona Maturation

The fix is separating “what I learned about the code” from “what I learned about how this person works.”

I introduced a harness-specific memory layer — an append-only ledger that lives inside the harness folder, not the project folder.

Two Streams of Memory

  1. Project trace (auditable): Stays in the repository’s session logs. Contains code-level details, dependency fixes, task history. Project B never sees Project A’s noise.
  2. Persona memory (epistemic): Stays in the harness folder (~/dotfiles/soul/harnesses/teddy-architect/v1/memory.jsonl). Contains distilled lessons about how I work and how the system should evolve.
[ SESSION ]
     |
     v
+----------------+      +-------------------+
| PROJECT TRACE  |      |  PERSONA MEMORY   |
| (Code Logs)    |      | (Harness Learning)|
+-------+--------+      +---------+---------+
        |                          |
  [ PROJECT A ]             [ GLOBAL HARNESS ]
  [ PROJECT B ]             [ PERSISTENT MIND ]

The Result

When the system boots, hydration is now triple-layered:

  • Static invariants: The base rules of the harness.
  • Persona maturation: Everything the harness has learned across all projects.
  • Project trajectory: The session logs of the current working directory.

The AI gets more precise and more aligned with my technical preferences every time I work with it. It’s not a temporary worker — it’s an accumulating asset.

Next: multi-agent specialization, where the harness spawns sub-agents for specialized reasoning domains.

← Back to main