dezgit2025/auto-memory

What it solves

AI coding agents often suffer from "context rot" or "amnesia" when context windows fill up and are compacted, or when new sessions start. This leads to a amount of time wasted re-explaining project details and files to the agent. auto-memory provides a way for agents to recall specific facts, files, and summaries from previous sessions without burning thousands of tokens on broad searches or manual re-orientation.

How it works

It acts as a read-only "page fault handler" for an agent's memory. Instead of keeping everything in the active context window (RAM), it pulls specific data from persistent storage (Disk) using a CLI tool.

• For GitHub Copilot CLI: It reads directly from the local SQLite database (session-store.db).
• For Claude Code: It reads JSONL session logs and builds an FTS5 index for structured search.
• Other Backends: It can also recall sessions from VS Code, JetBrains, and Neovim via file-backed providers.

The agent is instructed (via an instruction file) to run session-recall commands first to get oriented on recent files and session summaries before proceeding with a task.

Who it’s for

Developers using AI coding agents (specifically GitHub Copilot CLI, Claude Code, and IDE-integrated agents) who want to avoid the "death spiral of diminishing context" and reduce the time spent re-orienting their AI assistants.

Highlights

• Zero Dependencies: Built using only the Python standard library.
• Read-Only: Never writes to the underlying session databases or logs, ensuring safety.
• Token Efficient: Recalls specific context in ~50 tokens compared to thousands for manual grep/find operations.
• Multi-Backend Support: Works with Copilot CLI, Claude Code, and various IDEs.
• Agentic Installation: Includes protocols specifically designed for AI agents to install and configure the tool autonomously.