How I Built a Self-Improving Agent Architecture — A Loop Design Breakdown

Automatic skill generation, memory compression, and session archiving — how agents learn from experience

Summary

  • Implementing a "learn-from-experience" structure in AI agents yields progressively higher quality across sessions
  • A self-improvement loop that automatically generates skill candidates after complex tasks and patches problems immediately upon detection
  • Structuring and compressing memory to eliminate context waste
  • A session archive that extracts core knowledge even from deleted sessions

Background

Running AI agents surfaces a recurring set of structural problems: similar trial-and-error occurs each time a familiar task type appears, gains from one session do not transfer to the next, and performance degrades as context length grows.

Humans learn from experience and perform better over time. Can an agent operate on the same principle?

The OpenClaw project addressed this problem directly. OpenClaw is a multi-agent platform with hardened security and a stable self-improvement architecture. The design enables agents to automatically extract patterns from task experience, convert them into skills, and reorganize memory.

Body

1. Skill Self-Improvement Loop — The Agent Generates Its Own Skills

Conventionally, agent skills are authored by humans: write the skill file, specify the procedure, configure the trigger.

However, when an agent executes complex tasks, patterns emerge naturally. Combining multiple tools, or finding a resolution path through error recovery — that process itself is a skill candidate.

How the self-improvement loop operates:

Execute complex task (5+ tools used, error recovery included)
  ↓
Automatic analysis after task completion
  ↓
If repeatable pattern detected → register in skill-candidates.json
  ↓
Apply candidate skill in next task
  ↓
Patch immediately if problems are found

The critical constraint: a pattern must repeat 3 or more times before promotion to a skill. A single occurrence may be coincidence. Three repetitions signal structure.

Skill candidates are not pushed directly to production. They are applied experimentally in the next task. If a problem surfaces, it is patched immediately. The loop is: "This should work" → "This part breaks in practice" → "Fixed." Iterate.

2. Memory Compression — Core Context in 500 Tokens

An agent's context window is finite. As sessions lengthen, early instructions receive less weight and performance degrades.

Memory flush was restructured to solve this.

Previous approach: Agents stored memory in free-form text. No length limit. Tool outputs were often included in full. Memory grew bloated and wasted context.

Structured approach:

## Memory Flush
- Goal: [objective of this task]
- Progress: [completed / remaining]
- Decisions: [decisions made and their rationale]
- Changed Files: [list of modified files]
- Blockers: [open blockers]
- Next Steps: [next actions]

Three rules govern this: 1. Summarize tool output over 50 lines — discard the raw output. Re-run if the original is needed. 2. Target 500-token compression. A hard constraint that forces retention of only what matters. 3. Fix the schema to 6 fields. Free-form structures expand without bound.

The result: minimal tokens spent reading memory at session handoff, with core context preserved.

3. Session Archive — Learning from Deleted Sessions

When a session is deleted or reset in Claude Code, the conversation history disappears. If the insights gained in that session disappear with it, that is a net loss.

Session archiving solves this structurally.

How it operates:

Session end / deletion detected
  ↓
Run session-archive.py
  ↓
Auto-extract key content from .deleted / .reset sessions
  ↓
Store in staging/session-archive.json

What is extracted: - Problems solved and how they were solved - Decisions made and their rationale - Patterns discovered or cautions identified

What is not extracted: - Full source code (already in git) - Raw tool output (re-runnable) - Transient debugging traces

Sessions are ephemeral. What was learned persists.

4. From Observation to Verification — Dialectical Modeling

When an agent learns user behavior patterns, how does it validate what it has observed?

Drawing conclusions from a single observation is fragile. The user tag (U-tag) system applies a dialectical model.

3-stage verification:

Stage Condition Meaning
observation 1 occurrence "This tendency is visible"
hypothesis 2+ repetitions "This pattern is likely"
verified sustained over time, no counterevidence "Confirmed pattern"

A contradictions field tracks counterevidence. If a pattern is in verified state — say, "this user prefers data-driven decisions" — and an instance of intuition-based decision-making is observed, it is recorded as a contradiction and the status is reviewed.

Behavioral pattern assessments require the same validation rigor as data. This is a built-in defense against confirmation bias.

5. Automated Code Review — 3-Phase Separation

The process for agents reviewing agent-generated code was also redesigned.

Previous: Scan, analysis, and fix executed in a single pass. High false-positive rate. Fix scope unclear.

Redesigned: 3 distinct phases.

scan
  ↓ Generate candidate issue list
apply (analysis + queue)
  ↓ Filter to real issues only, enqueue for fix
fix (batch)
  → 2 files at a time, 10-minute cap

This enforces separation of generation and verification. Running both in a single pass degrades the quality of each. Separated, each operates correctly.

The batch fix limit of 2 files per 10 minutes exists because fixing too many files at once makes root-cause tracing difficult when something breaks.

Findings and Caveats

  • Skill auto-generation pitfall: The initial design promoted candidates after 2 repetitions. Coincidental patterns were being promoted to skills. Raising the threshold to 3 repetitions stabilized quality.

  • Over-aggressive memory compression loses context. At 300 tokens, "why this decision was made" was frequently dropped. 500 tokens is the current balance point.

  • Self-improvement can become an infinite loop. "Improving the skill that improves the skill that improves…" — meta-loops must be prevented. Self-improvement is restricted to trigger only after real task execution.

  • Lesson from the Hermes migration: During the early OpenClaw-to-Hermes transition, a token runaway event occurred. The self-improvement loop had been deployed to production without sufficient validation. The project reverted to OpenClaw and Hermes is currently being re-verified. The principle extracted from this: self-improvement structures must be fully validated in staging before production deployment.

Conclusion

Implementing a self-improvement architecture in an agent produces a system that compounds in quality with use. It stops repeating the same mistakes, extracts patterns from experience, and manages memory efficiently.

Three core principles:

  • Automate what repeats (3 occurrences → skill promotion)
  • Separate generation from verification (scan / analyze / fix — 3 phases)
  • Retain only the core, discard the rest (500-token compression, session archive)

Full autonomous evolution is a long way off. But the first step toward an agent that learns from experience is simpler than it seems: detect repeated patterns, record them, apply them next time. Not so different from how humans learn.

댓글

댓글 쓰기

이 블로그의 인기 게시물

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

이미지
← 8/9 Deep Learning Architec… 📚 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 »

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

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 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 (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 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. 핵심 요약 CNN : weight sharin...
Read more »

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

이미지
Series overview: Series index 📚 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. 핵심 요약 An evaluation set for RAG is ...
Read more »

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

← 6/9 Neural Networks 📚 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. 핵심 요약 SGD : \(w \leftarro...
Read more »

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

← 4/7 My Life Organizer Agent 📚 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 »

"ML Foundations (6/9) — Neural Networks: From Perceptron to MLP"

이미지
← 5/9 Regularization & Model… 📚 Series Index 7/9 Deep Learning Training → A neural network is logistic regression from Part 4 stacked into layers. Each layer passes through a new nonlinear activation, expanding the family of functions the network can represent. In this part we put how a neural network learns onto one page. Perceptrons, XOR, MLPs, backpropagation, activation functions, and the universal approximation theorem all meet here. 0. Learning Objectives Write the perceptron's learning rule and explain why it cannot solve XOR. Express an MLP's forward pass as one line of matrix products. Derive backpropagation by hand on a two-layer MLP. Compare the definitions, derivatives, and trade-offs of sigmoid, tanh, ReLU, and GELU. State the universal approximation theorem (Cybenko 1989 / Hornik 1991) precisely. Train an MLP end-to-end in PyTorch. 1. 핵심 요약 Perceptron (1958): \(\hat{y} = \mathrm{step}(w^\top x + b)\). Converges only on linearly separabl...
Read more »