gray and black laptop computer on surface

Persistent Memory for Claude: Five Approaches Compared

Claude does not remember anything between sessions on its own. Close Claude Code, open it tomorrow, and it has no idea what you decided about the database schema or which approach you already ruled out. Persistent memory for Claude is the set of tools and habits that fix that. There are now several, and they remember very different things. This guide compares five approaches so you can pick the right one, or the right combination.

The five approaches at a glance

ApproachWhat it remembersWhere it is storedWho decidesBest for
CLAUDE.mdRules, commands and conventions you writeA file in your repoYouEvery project, always
claude-memSummaries of past Claude Code sessionsLocally, SQLite plus embeddingsAutomaticDaily work on one codebase
ReporecallYour code: structure, call graphs, snippetsA local indexAutomaticLarge codebases
Private Journal MCPNotes Claude keeps about the workLocal, per project and globalClaudeWorkarounds and dead ends
MemSophYour corrections and preferencesHosted by MemSophYouChat, writing and research

1. CLAUDE.md: the memory you write yourself

A CLAUDE.md file at the root of a project is loaded at the start of every Claude Code session. It is the simplest and most reliable form of memory because nothing is guessed: Claude reads exactly what you wrote. Put your build and test commands, code style rules and the things Claude must never touch in it. Run /init to generate a starting file and /memory to edit it. There is more detail in /compact, /clear and CLAUDE.md explained.

The limit is size. The whole file is loaded every time, so a long one eats context in every session. Keep it to rules that change what Claude does.

2. claude-mem: automatic session memory

claude-mem hooks into the Claude Code session lifecycle. When a session ends it compacts the transcript into searchable memory, and the next session pulls the relevant parts back in. You do not have to write anything down, which is the main appeal. The trade-off is that retrieval is similarity search, so it can miss something phrased differently or bring back a decision you have since reversed.

3. Reporecall: memory of the code itself

Reporecall does not remember conversations at all. It indexes your codebase with tree-sitter and call-graph analysis, then injects relevant code context before Claude starts on a task. It saves Claude re-reading the same files at the start of every session, which matters most on big repositories. It works as Claude Code hooks or as an MCP server.

4. Private Journal MCP: notes Claude keeps for itself

The Private Journal MCP gives Claude somewhere to write things worth remembering that do not belong in code comments or commits: a workaround that worked, a preference you mentioned once, a dead end not worth revisiting. Entries are searchable by meaning, and there is both a project journal and a global one that follows you between repos. Claude decides what to write, so it is the least predictable of the five.

5. MemSoph: hosted memory you control

MemSoph is the odd one out. It is a separate AI chat app rather than a Claude Code add-on, and its memory is built from corrections you choose to save, which you can edit or delete later. It suits writing, research and general assistant work more than coding. Memory is stored by MemSoph rather than on your machine, so check its privacy policy before you put client work in it.

Memory and git worktrees

If you run several agents in parallel with git worktrees, each worktree starts a fresh session. CLAUDE.md is committed to the repo, so every worktree gets it automatically, which is one more reason to keep your core rules there. Tools that store memory outside the repo may tie it to a folder path, so check how they behave when you add a new worktree before relying on them.

Which should you use?

  • Start with CLAUDE.md. It is free and predictable, and every other option works better on top of it.
  • Add claude-mem if you work on the same codebase every day and keep re-explaining past decisions.
  • Add Reporecall if Claude spends the start of each session re-reading a large codebase.
  • Try the Private Journal MCP if you want Claude to keep track of workarounds and dead ends itself.
  • Look at MemSoph if your AI work happens mostly in a chat window rather than a coding agent.

Whatever you choose, prune it. Stale memory is worse than none, because it confidently pushes Claude towards decisions you have already changed. The same idea runs through context engineering for AI agents: what you feed the model matters more than how much.


Leave a Reply