Agent Self-Improvement Harness (1/12) — Separating Static Identity from Runtime SOUL

The Problem — Two Files with the Same Name, 47% Drift

AI agent "persona" data splits into two layers. A static identity layer that rarely changes (values, worldview, capability definitions) and a dynamic state layer updated every session (learning notes, observations, user insights). When both layers share the same file, editorial authorities compete and drift accumulates.

In one actual agent, the SOUL identity file existed in both the repo (persona/) and the runtime (hermes-data/) directories:

  • Initial template (repo, 16 KB) versus actual runtime file (24 KB, updated with mir self-edits)
  • A 47% difference between the two files
  • USER.md was worse — the repo version held a static profile, while the runtime version had accumulated §-separated insights and evolved into a completely different file

The root cause was storing two editing authorities in one path. When a manager edits manually and the agent self-edits on the same path, one side will always lag or get overwritten.

The Design — Split Paths by Editing Authority (Option C Hybrid)

The fix is to separate the two layers explicitly and give each layer exactly one authoritative path.

Static identity — repo as single source

Path Role Author
persona/ONTOLOGY.md 4-layer identity framework Manager
persona/VALUES.md Value priorities, judgment rules Manager
persona/BEING.md Temperament, worldview, non-negotiable principles Manager
persona/CAPABILITIES.md Skills / meta-skills / knowledge domains Manager
persona/LIFE_PROJECT_PLAN.md Project direction Manager

Dynamic slots — runtime as single source

Path Role Author
hermes-data/SOUL.md Runtime-injected SOUL Manager + mir self-edit
hermes-data/memories/USER.md Accumulated user insights Agent (memory tool)
hermes-data/memories/MEMORY.md Session memory Agent (memory tool)

Historical files (persona/SOUL.md, persona/USER.md) are deprecated. The persona/loader.sh script that handled copying is deprecated too, retaining only the --uninstall path.

The Implementation — Idempotent One-Way Ingest

The path from static identity to memcore (a SQLite-based curated memory) collapses into one script.

persona/*.md (repo single source)
       │
       ▼  scripts/persona-ingest.py  (manually invoked by manager)
       │
       ▼
memcore curated
  source_path = "persona/<file>.md"

Operating principles:

  • Direction: repo → memcore one-way. No back-flow.
  • Idempotence: DELETE+INSERT pattern. On re-run, prior N rows → inserted N rows, unchanged.
  • Traceability: source_path records the unique path. Reverse-traceable.
  • Chunk-row match: H2 section count in the file equals curated row count. File structure maps 1:1 to storage structure.

No automatic hook is attached. Persona changes happen a handful of times per year at most — there is no reason to ingest on every boot, and the editing cadence must be decided by the manager.

Verification — Six-Axis Checklist

File separation alone is not structural governance. Separation only matters if these six axes all pass.

  1. Idempotence — dry-run / run / re-run yield the same row count
  2. Chunk-row match — H2 section count equals curated count
  3. Traceability — unique source_path values match the file count
  4. features-sync integrity — existing entries in the memcore wiki remain unchanged (no drift)
  5. Wiki reflection — feature document sections stay synced with curated state
  6. Rollback path — uninstall script restores the historical-file setup

Without all six, "separating two files" is indistinguishable from "changing nothing structural."

The Effect — The Drift Path Itself Disappears

After separation, the paths that could produce drift are gone. Static identity flows one way from repo to memcore, and the runtime SOUL is edited directly under hermes-data/. The two paths never cross, so the question "which copy is current?" no longer exists.

Side effect: observation points also become clear. If the runtime SOUL exceeds its token budget (~8.3 K tokens against a 3 K recommendation — 2.7× over), you monitor hermes-data/ alone. The repo side's change history lives in git log. When responsibility splits, monitoring targets split with it.

Application — A Generalizable Pattern

This separation pattern generalizes to "any config file where manager edits and agent self-edits intermix."

  • System prompts + learning notes → templates/ (repo) + runtime/notes.md (runtime)
  • Tool schemas + tool usage stats → tools/schemas/ (repo) + runtime/usage-log.db (runtime)
  • Agent role definitions + role execution records → roles/ (repo) + runtime/role-history.md (runtime)

The shared rule is one sentence: never put two editing authorities on one file. When editors differ, put the files in different places, and make the flow one-directional — and drift itself stops occurring.

Conclusion

File separation looks trivial, but for structures where editing authorities were already mixed, it is a decisive shift. The cost of detecting and resolving drift disappears, and a problem on one side no longer propagates to the other. The "single source" principle is not mere documentation hygiene — it is a structural isolation mechanism. Separation only functions as real governance when paired with idempotent one-way ingest and six-axis verification.

References

  • persona/module.yaml v2.0 — hybrid rules + deprecated section
  • scripts/persona-ingest.py — idempotent ingest implementation
  • docs/operations/features/persona-loader.md — feature spec
  • docs/operations/persona-loader-verification-2026-04-21.md — six-axis verification report

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »