Project Memory: A No-Install Alternative to claude-mem
A no-install alternative to claude-mem: your agent keeps decisions, gotchas and session notes as markdown in .agent-memory/, with no hooks, servers or database.
Updated November 2026.
Agent memory
Plain markdown
No install
Cross-agent
Secret scrubbing
About This Skill
Project Memory keeps decisions, gotchas, conventions and unfinished work as plain markdown in your repo, and the agent reads a short index first and opens only the notes that match the task. Where claude-mem uses lifecycle hooks, a local worker service and a SQLite database, this skill needs nothing installed or running, and the same folder works across Antigravity, Claude Code and Codex. Inspired by Claude-Mem by Alex Newman (Apache-2.0). Not affiliated.
New to the original tool? Read our claude-mem guide: what it does, how to install it and how this skill differs.
Quick Start
1Download the skill (1 credit) and unzip the project-memory folder.
2Copy the folder into your skills directory: Antigravity ~/.gemini/config/skills/ (or .agents/skills/ in a project), Claude Code ~/.claude/skills/, Codex ~/.agents/skills/.
3Ask the agent to set up project memory; it asks once before creating .agent-memory/ and whether to commit or git-ignore it.
4Work as usual: the agent recalls matching notes at the start of non-trivial tasks and records new decisions and gotchas at the end.
Example Command
Set up project memory for this repo
Core Capabilities
Plain markdown files
An index plus decisions, gotchas, conventions and a session log, kept in .agent-memory/ at the repo root.
Index-first recall
Reads INDEX.md, opens only the notes that match the task, and checks recalled facts against the current code.
Scrubbed writes
Never stores secrets, personal data or anything inside <private> tags; refers to credentials by name only.
Deduplication
Updates existing entries instead of adding copies, and keeps a short superseded line when a decision changes.
Pruning and compaction
Removes stale notes, merges duplicates, compacts the session log and rebuilds the index when limits are reached.
Works across agents
Antigravity, Claude Code, Codex and other agents share one folder; each entry records its source.
Usage Examples
Input
What did we decide about background job locking?
Output
The agent reads INDEX.md, opens the matching entry in decisions.md, verifies the referenced code still exists, and answers with the decision and its date.
Input
Remember that the staging seed script must run before e2e tests.
Output
A short dated entry in conventions.md or gotchas.md, a matching INDEX.md line, and a one-line note saying what memory changed.
Input
Forget everything about the old payment provider.
Output
The agent lists the entries it will remove, waits for confirmation, deletes them and their index lines, and reminds you that committed text remains in git history.
SKILL.md
---
name: project-memory
description: Keeps a persistent, file-based project memory in plain markdown inside the repo (`.agent-memory/`) so the agent remembers decisions, gotchas and past sessions without hooks, servers or databases. Use at the start of any non-trivial task in a repo that has `.agent-memory/`, when the user asks to "remember", "recall", "what did we decide", "set up project memory", or at the end of a task that produced a decision, a non-obvious fix or a lesson worth keeping.
---
# Project Memory
Persistent memory for coding agents, stored as plain markdown in the repository.
No install, no background process, no database, no network. Any agent that can read
and write files (Google Antigravity, Claude Code, Codex, and others) uses the same
folder the same way, so memory written by one agent is readable by the next.
The model is simple:
1. **Read small first.** Open the index, never the whole folder.
2. **Open only what matches.** Follow index entries to individual notes.
3. **Write little, write late.** At the end of a task, record only what a future
session would otherwise have to rediscover.
4. **Keep it clean.** Merge duplicates, delete stale notes, never store secrets.
## When to use
- **Start of task:** the repo has `.agent-memory/INDEX.md` and the task is more than a
one-line change. Run "Recall".
- **End of task:** the work produced a decision, a non-obvious cause, a workaround,
a convention, or unfinished work. Run "Record".
- **User asks:** "remember this", "what did we decide about X", "forget X",
"set up memory", "clean up memory". Run the matching section.
- **Skip it** for trivial edits, pure questions about general knowledge, or when the
user says not to use memory for this task.
## Folder layout
```
.agent-memory/
INDEX.md one line per note; the only file read every time
decisions.md architecture and product decisions, with the why
gotchas.md traps, surprising behaviour, non-obvious fixes
Frequently Asked Questions
FAQ
How is this different from claude-mem?
Claude-Mem relies on agent lifecycle hooks, a local worker service and a SQLite database with optional vector search. Project Memory replaces all of that with markdown files the agent reads and writes itself, so there is nothing to install or run and memory can be committed or git-ignored.
Is this the official claude-mem?
No. It is not Claude-Mem and is not a port, fork or plugin of it, and it is not affiliated with Alex Newman. The official project is at https://github.com/thedotmack/claude-mem (Apache-2.0).
Which agents can use it?
Any agent that can read and write files, including Google Antigravity, Claude Code and Codex. Memory written by one agent is readable by the next.
Should the memory folder be committed?
Your choice. Committing shares memory with the team and reviews it like code; ignoring keeps it personal via .gitignore or .git/info/exclude. The skill asks before creating the folder.
What does a download cost?
Each skill download costs 1 credit. New accounts start with 3 free credits.