Agent Memory Engine (8/10) — Applying Karpathy LLM Wiki in Production

Implementation record: 17 of Andrej Karpathy's 18 proposed features implemented, plus 6 custom additions

Summary

  • Applied Karpathy's LLM Wiki concept in full to the OpenClaw agent platform
  • 17 of the original 18 features are implemented; 6 additional features address gaps found in production
  • Instead of searching external documents on every query (RAG), the LLM incrementally builds and maintains a permanent, structured wiki

What Is an LLM Wiki

The core idea Andrej Karpathy published is straightforward. Rather than retrieving external documents on every query (RAG), the LLM itself incrementally builds and maintains a structured markdown wiki.

The wiki operates through three actions:

  • Ingest: Accepts raw sources and updates wiki pages. Triggers cascade updates to related pages.
  • Query: Answers questions using the wiki. Knowledge gaps discovered during answering are promoted into the wiki.
  • Lint: Audits the entire wiki. Detects and resolves contradictions, gaps, and stale entries.

When these three actions form a loop, the knowledge base self-refines without human intervention.

Body

1. Karpathy's Original 18 Features vs. OpenClaw Implementation

# Karpathy Feature OpenClaw Implementation Status
1 Raw Sources (immutable originals) memory/ + docs/manuals/ ✅
2 Wiki (LLM-maintained markdown) bank/ multi-file Context Tree ✅
3 Schema (co-evolving config) tree-index.json + AGENTS.md ✅
4 Ingest cascade updates retain-merge cascade_update ✅
5 Interactive Ingest source-ingest.py --interactive ✅
6 Automated external source Ingest source-ingest.py --url/--file ✅
7 Query (wiki-backed answers) recall-tree.py BM25 hybrid ✅
8 Query → Wiki promotion recall-tree.py --promote ✅
9 LLM reranking recall-tree.py --rerank (oMLX gemma-4) ✅
10 Basic Lint bank-lint + memory-warning-report ✅
11 Lint exploration suggestions exploration.page_gap_suggestions ✅
12 Lint unresolved contradiction tracking exploration.unresolved_contradictions ✅
13 index.md (catalog) tree-index.json (full file index) ✅
14 log.md (chronological record) _meta/changelog.md ✅
15 Git version control git tracked ✅
16 Schema co-evolution self-review cron auto-update ✅
17 Batch vs. interactive toggle default=1 item, Reflect=batch ✅
18 Obsidian graph view N/A (CLI environment) —

17/18 implemented. The only unimplemented feature is the Obsidian graph view, which is not applicable in a CLI environment.

2. Ingest / Query / Lint: Concrete Implementation

Each of the three actions is described in detail below.

Ingest


def ingest(source, interactive=False):
    content = fetch(source)
    tags = TAG_ROUTING.classify(content)   # Deterministic regex-based classification
    target_files = resolve_targets(tags)   # W/B/O/S → resolve target files
    for f in target_files:
        cascade_update(f, content)         # Cascade update
    changelog.append(source, tags)

TAG_ROUTING classifies content into four tag types — W (world facts) / B (activities) / O (opinions) / S (status) — using regex keyword rules. The decision to use deterministic rules instead of LLM judgment is explained in the lessons-learned section below.

Query


def query(q, rerank=False, promote=False):
    candidates = bm25_hybrid_search(q)    # BM25 + keyword hybrid
    if rerank:
        candidates = llm_rerank(candidates, model="oMLX/gemma-4")
    answer = generate_answer(q, candidates)
    if promote and has_knowledge_gap(answer):
        wiki.create_stub(q)               # Gap → create new wiki stub page
    return answer

Korean agglutinative morphology handling is the key challenge. "๊ฐ€์น˜๊ด€์ด" → "๊ฐ€์น˜๊ด€" morpheme normalization is applied at the BM25 indexing stage. English-centric search libraries do not provide this out of the box.

Lint


lint_checks = [
    check_contradictions,      # Detect unresolved contradictions
    check_page_gaps,           # Detect linked-but-missing pages
    check_stale_entries,       # Entries below confidence-decay threshold
    check_orphan_pages,        # Pages referenced by nothing
    check_data_gaps,           # Entries with figures but no source
]

Lint runs in two modes: manual and automated. The Reflect cron (03:00) executes a full batch Lint; the micro-cycle (every 30 minutes) runs a lightweight version.

3. Six OpenClaw-Specific Features Not in Karpathy's Proposal

These features address problems discovered in production that the original proposal does not cover.

1. confidence-decay (Opinion Confidence Decay)

Opinion files are assigned a confidence score that auto-decrements by −0.02 per day. Entries falling below 0.30 are automatically removed. Karpathy mentions contradiction detection, but quantifying the temporal value of opinions is a separate problem. Old opinions and recent opinions must not carry equal weight.

2. fuzzy dedup (Fuzzy Duplicate Prevention)

Duplicate accumulation is blocked at an 80% keyword overlap threshold. "Flutter state management is good" and "Flutter์˜ state management๊ฐ€ ์ข‹๋‹ค" are semantically identical despite different surface forms — exact-match deduplication cannot catch this.

3. TAG_ROUTING (Deterministic Tag Routing)

W/B/O/S tags are auto-classified via regex keywords to determine target files. "Tool change" → world/tools.md; "Project" → world/projects.md. Deterministic rules are chosen over LLM judgment to ensure routing stability.

4. proactive briefing (Autonomous Briefing)

A multi-mode proactive briefing system that surfaces relevant wiki content without being asked. Includes a conversation-context-aware mode. The wiki should not only respond to manual searches — it must proactively surface information.

5. Three-Tier Automated Distillation Pipeline

memory/ (daily log) → bank/ (curated knowledge) → recall/ (search buffer) — three stages running automatically. Reflect (full distillation at 03:00 daily) and micro-cycle (lightweight distillation every 30 minutes) refine knowledge without human intervention.

6. Korean Agglutinative BM25 Partial Matching

"๊ฐ€์น˜๊ด€์ด" → "๊ฐ€์น˜๊ด€" morpheme normalization. Not provided out of the box by English-centric search systems.

4. RAG vs. Wiki Paradigm

RAG approach: query → vector search → top-N chunk retrieval → LLM answer - Advantage: simple to set up - Disadvantage: answer quality degrades on retrieval failure; knowledge does not accumulate

Wiki approach: LLM writes knowledge directly into structured markdown → search traverses structured files - Advantage: knowledge is progressively refined; structured files yield higher retrieval precision - Disadvantage: initial build requires effort; maintenance pipeline required

The wiki approach is decisively superior in long-term operation because of cumulative knowledge refinement. RAG searches raw chunks on every query; the wiki searches the refined structure built up by prior Ingest, Query, and Lint passes. Answer quality for the same query improves as the system matures.

5. Measured Results

Metric Value
Search precision 10/10 natural language queries
File size Peak 46 KB → 10.6 KB (75% reduction)
Duplicate accumulation 0 after fuzzy dedup
Topic coverage 100% across major topics
memory-warning count Many → few
LLM accuracy (Gemma 4) 16/16 (100%)

Lessons Learned

Wiki was not designed from the start: Knowledge was accumulated in a single file until it bloated to 46 KB before the need for structure became apparent. The schema had to be designed using Karpathy's framework as a reference, then the existing file was decomposed and reclassified in reverse.

Overreliance on LLM judgment: Initial tag classification was delegated to the LLM, but nondeterministic outputs caused the same input to route to different files. Stability was restored after introducing TAG_ROUTING with deterministic rules.

Lesson from the OpenClaw → Hermes migration: A migration to Hermes was attempted while OpenClaw was in stable operation, but token runaway occurred. After reverting to OpenClaw, Hermes redesign is currently under validation. This process confirmed that the wiki system must remain platform-independent.

Conclusion

Karpathy's LLM Wiki operates as a loop of three actions: Ingest, Query, and Lint. Applying this structure in production achieves a level of knowledge refinement and retrieval precision unreachable with the RAG approach. The key is not copying the idea verbatim, but redesigning it for your own environment — language, operational model, and data characteristics.

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 »