Agent Memories

A portable, project-agnostic memory and skill scaffold for AI coding agents (Claude Code, Gemini/Antigravity, and similar).

The core idea: an agent that works across many different codebases accumulates two kinds of knowledge — knowledge that is useful everywhere (how you like to communicate, how you like commits done, reusable tool workflows) and knowledge that only makes sense for one specific project (its language, its build system, its domain quirks). Mixing the two makes the "everywhere" knowledge unusable when you switch to a different project. This repository keeps only the first kind. Project-specific knowledge lives inside each project's own repository instead.

Structure

Agent_Memories/
├── Claude_Memory/          # Claude-specific working memory
│   ├── MEMORY.md            # index — start here
│   ├── claude_info/         # CHECKPOINT / NOTES / TIPS / TODO (session-scoped scratch files)
│   ├── claude_log/          # dated work logs (LOG_YYYYMMDD_<topic>.md), format only — see README inside
│   ├── claude_skills/       # mirrored Skill definitions (archive-session, graphify, grill-me, handoff)
│   ├── feedback_*.md        # behavioral rules learned from user corrections/confirmations
│   ├── user_*.md            # stable facts about the user (preferences, profile)
│   └── reference_*.md       # reusable tool/reference documentation (e.g. graphify usage)
├── caveman/                 # ultra-compressed communication mode (skill + subcommands)
├── html-slide-deck/         # skill for building self-contained HTML slide decks
└── html-slide-deck.skill    # packaged version of the above

Memory file conventions (`Claude_Memory/`)

Prefix	Meaning	Changes how often?
`feedback_*.md`	A behavioral rule learned from an explicit user correction or confirmation. Each file states the rule, why it exists, and how to apply it.	Rare — only when the user corrects or confirms an approach
`user_*.md`	Stable facts about the user: preferences, working style, profile.	Rare
`reference_*.md`	Documentation for a reusable tool or workflow that isn't tied to any one project.	Occasional
`claude_info/*.md`	Short-lived session scratch space (current status, discoveries, todos).	Every session
`claude_log/*.md`	Dated logs of non-obvious discoveries, kept for future reference.	Per notable discovery

Rule of thumb: if a note only makes sense to someone who already knows the specific project you were working in, it does not belong here — move it into that project's own memory/notes instead. This repository was extracted from a working setup that originally mixed both kinds of content; project-specific material was moved out during that cleanup.

Usage

Clone or copy this repository's contents into wherever your agent looks for persistent memory (for Claude Code / Cowork-style setups, that's typically a fixed path referenced from your global CLAUDE.md).
Point your agent's session-start instructions at Claude_Memory/MEMORY.md — it's the index and links to everything else.
As you work, let the agent keep claude_info/ and claude_log/ current. When a note turns out to be project-specific, move it into that project's own memory instead of leaving it here.
Skills under caveman/ and html-slide-deck/ can be installed/loaded independently of the memory convention above; they're self-contained.

Notes

graphify-out/ (a generated knowledge-graph cache used by the graphify skill) is intentionally git-ignored — it's regenerated per-project on demand and shouldn't be committed.
This repo assumes no particular OS, language, or toolchain. If you find something in here that only makes sense for one specific stack, that's a bug — please move it out.

Commit Convention

Commit messages in this repository follow the Angular commit message convention:

<type>(<scope>): <subject>
<BLANK LINE>
<body>
<BLANK LINE>
<footer>

type: feat, fix, docs, style, refactor, perf, test, chore
scope: optional, whatever area of the repo the change touches (e.g. feedback, caveman, readme)
subject: imperative, present tense, no capital first letter, no trailing period
Lines wrapped at 100 characters
Reverts start with revert: followed by the original commit's header, with This reverts commit <hash>. in the body
Breaking changes are called out with a BREAKING CHANGE: line in the footer

License

MIT — see LICENSE.

4.7 KiB Raw Blame History