← marketplace
engineerstoolsha:fdcdb68345cc91bbmanual

claude-mem

Use when you want Claude Code to remember past sessions — it compresses each session's work into observations and auto-injects the relevant ones so you stop re-explaining your codebase.

Install confidence
curl --create-dirs -fsSL https://skillmake.xyz/i/claude-mem -o ~/.claude/skills/claude-mem/SKILL.md
Pinned content
sha:fdcdb68345cc91bb
Generated with
manual
Source
github.com

The file served at /api/marketplace/claude-mem-fdcdb683/raw matches this hash. Inspect before install, then copy the command.

3,046 chars · ~762 tokens
---
name: claude-mem
description: "Use when you want Claude Code to remember past sessions — it compresses each session's work into observations and auto-injects the relevant ones so you stop re-explaining your codebase."
source: https://github.com/thedotmack/claude-mem
generated: 2026-07-23T01:35:04.192Z
category: tool
audience: engineers
---

## When to use

- You keep re-explaining the same codebase context at the start of every session
- You want to ask 'did we already solve this?' or 'how did we do X last time?' across sessions
- You want decisions and discoveries from prior work carried forward automatically
- You want a searchable, local knowledge graph of everything an agent has done in a project

## Key concepts

### Observations from tool calls

Every Read, Edit, and Bash the agent runs is compressed into an observation. Observations are summarized at session end, and relevant ones are auto-injected into future prompts.

### Injection starts on session two

The first session in a fresh project seeds memory; from the second session onward the plugin auto-injects context for relevant past work so the agent starts already oriented.

### Local-first storage

Everything lives in ~/.claude-mem on your machine — a SQLite database, vector index, logs, and settings. Nothing leaves except calls to whichever provider you configured for compression.

### Memory search workflow

The mem-search skill follows search -> filter -> fetch to pull work from previous sessions, answering questions about what was done before rather than the current conversation.

### learn-codebase front-load

Running /learn-codebase does an optional one-time (~5 minute) pass that loads the entire repo into memory up front instead of accumulating context session by session.

## API reference

```
claude plugin marketplace add thedotmack/claude-mem
```

Add the claude-mem marketplace so the plugin can be installed.

```
claude plugin marketplace add thedotmack/claude-mem
```

```
claude plugin install claude-mem@thedotmack
```

Install the claude-mem plugin (memory hooks + MCP search server).

```
claude plugin install claude-mem@thedotmack
```

```
/learn-codebase
```

Optional one-pass front-load of the whole repository into memory (~5 minutes).

```
/learn-codebase
```

```
npx claude-mem uninstall
```

Cleanly remove claude-mem and delete its ~/.claude-mem data directory.

```
npx claude-mem uninstall
```

## Gotchas

- Memory injection only kicks in from your second session in a project — the first session just seeds it
- Compression sends observations to whichever AI provider you configure (Claude / OpenRouter / Gemini), so it is not fully offline
- State is per-machine under ~/.claude-mem — it does not sync across machines unless you enable cloud-sync
- mem-search answers about PREVIOUS sessions, not the current conversation — use it for 'how did we do X last time?'

---
Generated by SkillMake from https://github.com/thedotmack/claude-mem on 2026-07-23T01:35:04.192Z.
Verify against source before relying on details.

File: ~/.claude/skills/claude-mem/SKILL.md