llmwiki: Solving the LLM Context-Switch Tax with Persistent Project Memory

I Got Tired of Re-explaining My Codebase to Claude Code Every Session. So I Built llmwiki.

Engineer Max Małecki developed llmwiki to automate the creation of persistent, version-controlled project documentation for AI coding assistants. The tool leverages a materialize command to rebuild context from accumulated facts, cutting token usage from 100K down to 15K per session.

Why This Matters

LLM coding assistants like Claude Code often lack long-term memory, forcing developers to expend high token counts or manual effort re-explaining architecture in every new session. llmwiki addresses this technical reality by decoupling project knowledge from ephemeral chat contexts, providing a structured Markdown layer that syncs via Git and integrates with tools like Obsidian for visual mapping. This approach mitigates the cognitive tax of context-switching between projects while significantly lowering the operational cost of using high-end models like Opus 4.7.

Key Insights

• Token cost reduction of ~85% for project context (15K vs 100K tokens) using the materialize command in 2026.
• Persistent Markdown wiki pattern popularized by Andrej Karpathy to maintain project knowledge across session boundaries.
• Integration of Graymatter memory layer for semantic search and fact decay with a 30-day half-life to prune stale information.
• Support for local-first execution using Ollama backend to ensure data privacy and security for air-gapped environments.
• Automated documentation generation including Mermaid diagrams, C4 system landscapes, and OpenAPI-extracted API docs.

Working Examples

Scans the project to generate a markdown wiki covering architecture, service maps, and API docs.

llmwiki ingest ~/workspace/my-api

Rebuilds the wiki from accumulated facts, reducing token usage compared to a full ingest.

llmwiki materialize my-project

Injects the project map into a CLAUDE.md file using marker blocks.

llmwiki context my-project --inject CLAUDE.md

Marker blocks used for automated context injection in documentation files.

<!-- llmwiki:start -->
... domain, architecture, services, flows ...
<!-- llmwiki:end -->

Practical Applications

• Use case: Tech leads onboarding junior engineers use auto-generated C4 system landscape diagrams to visualize service relationships across multiple repos. Pitfall: Relying on manually updated Confluence pages which are often outdated by years.
• Use case: Consultants managing multiple client stacks utilize the Graymatter layer for semantic search across project-specific facts. Pitfall: Manually re-explaining architecture to AI assistants every morning, leading to wasted billable hours and high token invoices.
• Use case: Security-sensitive environments run llmwiki with an Ollama backend for air-gapped codebase analysis. Pitfall: Sending proprietary source code to cloud-based LLM providers without proper path-traversal or data-leakage safeguards.

References:

• https://github.com/emgiezet/llmwiki
• https://dev.to/emgiezet/i-got-tired-of-re-explaining-my-codebase-to-claude-code-every-session-so-i-built-llmwiki-4j2k