Claude Memory Compiler

Why it's ranked here

The design is unusually legible and inspectable. Markdown, source links, build logs, and health reports make the stored memory easier to audit than an opaque retrieval service. Automatic capture also reduces upkeep. The catch is substantial dependence on Claude Code and model judgment, while index-guided retrieval deliberately targets personal rather than large knowledge collections.

What's good

The source conversations remain separate from compiled articles, preserving a trail back to the original context. Changed logs are detected with hashes instead of being blindly reprocessed. Capture runs in background processes, with recursion prevention and duplicate-flush protection. The health checker catches broken links, uncompiled sources, stale content, missing backlinks, sparse articles, and possible contradictions.

Tradeoffs

Extraction, compilation, contradiction checks, and question answering depend on model output, so article quality and consistency are not guaranteed by the structural checks. Session capture keeps only recent turns and imposes character limits, which can omit earlier detail. Startup context is also truncated. Retrieval reads the assembled wiki content, and the documentation says a larger collection eventually needs RAG.

How to use it well

Use it for a long-running Claude Code project where decisions, debugging lessons, and recurring conventions otherwise disappear between sessions. Keep daily logs as the source record, compile regularly, inspect health reports, and file back answers that deserve reuse. It does not serve as a general document ingestion system, and it is not the proposed retrieval approach once the collection outgrows an index that fits in context.

Observed

Primary language

Python

Runtime requirement

Python 3.12 or newer

Installation surface

uv project with a committed lockfile

Dependencies

Claude Agent SDK, python-dotenv, and tzdata

Interfaces

Claude Code hooks and command-line scripts

Storage format

Local Markdown articles, daily logs, JSON state, and Obsidian-style wikilinks

Retrieval approach

Index-guided LLM selection without embeddings or a vector database

Configured hook events

SessionStart, PreCompact, and SessionEnd