agentlog

agentlog is a Python CLI by Ryan Alberts that keeps a persistent decision log for any project. Decisions are stored as JSONL lines, and Claude or Gemini is used to retrieve relevant decisions and surface patterns across them. It is published on PyPI as agentlog-cli (version 0.1.0, released April 17, 2026) under the MIT license.

What It Is

agentlog is a decision-memory tool for teams and individuals who want to keep the reasoning behind choices instead of losing it in chat threads and documents. The README describes it as Karpathy's LLM Wiki idea shipped as a CLI with three commands: remember, recall, and reflect.

How the workflow works

remember stores a decision as a JSON line with an id, timestamp, text, and auto-generated tags. Text can be typed directly or produced by piping a summary from another LLM CLI. recall sends a query plus all stored decisions to Claude or Gemini and returns the relevant context with connections explained. reflect asks the model to find patterns, contradictions, and gaps across the whole history, and saves the results to .agentlog/reflections.md so they accumulate.

Design choices and model access

The project documents a deliberately minimal design: one file of roughly 200 lines of Python, JSONL storage that is grep- and git-friendly, and no embeddings, vector database, or RAG pipeline. The README states this approach works up to about 500 decisions before chunking would be needed. Claude is the default provider; Gemini can be selected per command with --model gemini or by setting AGENTLOG_PROVIDER.

Setup path

Install with pip install agentlog-cli and set ANTHROPIC_API_KEY or GEMINI_API_KEY. Both provider SDKs are installed as dependencies. The package requires Python 3.9 or newer, and the CLI binary is named agentlog. It can also be run from source by cloning the repository and using pip install -e ..