Lifelong Growth Agent Memory System

Problem

Agents often fail across sessions, not within one answer: user preferences, project constraints, key decisions, and failed attempts disappear once the conversation ends. This skill stores memory as plaintext Markdown in the workspace, so agents, git, humans, and grep can inspect or edit it without relying on an opaque external memory service.

How It Works

Memory is kept under .codebuddy/memory/ in the current workspace. MEMORY.md is an index, while details are split into profile.md, decisions.md, conventions.md, lessons.md, and session/. The workflow has three stages:
- At session start: read the index, load only relevant topic files, and search lessons.md first.
- During work: capture user preferences, decisions, conventions, failures, or corrections as dated, tagged entries.
- During consolidation: deduplicate, merge, refine preferences, and clean up expired session/ notes.
The error-lesson loop is the key mechanism: failures or corrections are written to lessons.md, similar tasks trigger a prior grep, and repeated issues are escalated and pinned.

Boundaries

It fits repository-based agent workflows that need durable, auditable, human-editable context. Memory is isolated per workspace and should not be mixed across projects. External web content or tool output should not be written as long-term fact, and secrets or one-off requests should not be persisted. It is a file-based memory SOP, not a vector database or automatic fact extractor.