Compare

Wordcell and Claude-Mem

Claude-Mem records what your coding agent does in each session, has a model condense it, and loads recent observations into the next session without being asked. Wordcell keeps only the notes you or your agent choose to write, as Markdown files in your repository. Pick Claude-Mem for memory that builds itself; pick Wordcell for memory you review in Git and get with every clone.

Claude-Mem 13.30.0, its documentation, and its source were checked on October 4, 2026.

How they differ

Wordcell and Claude-Mem at a glance
	Wordcell	Claude-Mem
Writes memory without being asked	No	Yes
Loaded at session start	No notes until the agent searches	50 recent observations by default
Stored as	Markdown files	SQLite and a vector index
Kept in	Your repository or a folder you choose	~/.claude-mem on one machine
After a fresh clone	The notes come with it	On the same machineSame folder name by default
Model calls to save memory	None	Every captured session
Codex can search Claude Code’s memory	Yes	YesSession start needs a setting
Software cost	Free · MIT	Free · Apache-2.0

Claude-Mem’s cells come from its documentation and its 13.30.0 source, checked October 4, 2026; it was not run for this page. Its memory model runs on your Claude plan, your Codex sign-in, your own key, a local model server, or its paid CMEM Pro service. Another machine gets Claude-Mem’s database only through an export or cloud sync.

Sources and details

How memory forms: Claude-Mem: Its hooks run on every session start, prompt, and tool call. A model condenses the captured tool calls into observations with a title, narrative, and facts, and writes a summary when Claude finishes responding. Content inside <private> tags is not stored.; Wordcell: You or your agent write notes, with wordcell note create, the create_note tool, or any editor. Wordcell writes no note on its own.
Session start: Claude-Mem: The SessionStart hook adds the project’s recent observations, 50 by default, to the agent’s context as a timeline of titles, with the latest session summary when it is newer than those observations.; Wordcell: No notes are loaded. The agent calls search, or context for the file it is changing, when it needs memory.
Storage: Claude-Mem: A SQLite database at ~/.claude-mem/claude-mem.db, a Chroma vector index, settings, and logs, all under ~/.claude-mem.; Wordcell: Markdown files with YAML frontmatter in a folder you choose, usually kb/ in the repository.
Fresh clone or another machine: Claude-Mem: Memory is filed under the repository folder’s name, or under its org/repo remote with CLAUDE_MEM_PROJECT_NAME_SOURCE=git-remote. Another machine needs an export and import or cloud sync, which uploads observations, summaries, and prompts to cmem.ai. Its opt-in folder CLAUDE.md files, off by default, write a timeline of recent activity into the repository’s subfolders.; Wordcell: A committed vault arrives with every clone, and Git history shows who changed each note.
Model use: Claude-Mem: A model processes every captured session: your Claude plan, your own OpenRouter or Gemini key, your Codex sign-in, an OpenAI-compatible endpoint such as a local Ollama or LM Studio server, or the CMEM Pro service that the interactive installer selects first and offers free for up to 14 days.; Wordcell: Writing notes and exact search need no model. Semantic search runs a local model after a one-time download, and hosted reranking is optional and paid.
Agents: Claude-Mem: Installers for Claude Code, Codex CLI, Cursor, Windsurf, OpenCode, and others, and MCP tools that search its memory. A search covers every agent’s observations unless the call names one platform. Codex’s session start includes Claude Code’s observations only with CLAUDE_MEM_SESSION_START_INCLUDE_ALL_SOURCES set.; Wordcell: Any MCP client that starts a local server, including Claude Code, Codex, and Cursor; see Connect Wordcell to your coding agent. The CLI works without MCP.
Review and edit: Claude-Mem: A local web viewer shows the memory stream. Observations are rows in the database.; Wordcell: Edit notes in any editor and review each change with git diff.
License: Claude-Mem: Apache-2.0.; Wordcell: MIT.

Where the text goes

The two tools keep your work in different places, and only Claude-Mem sends it to a model to save it.

Claude-Mem

Hooks in Claude Code, or in another agent it supports, send each tool call and its result to a local worker on your machine.
The worker sends them to the memory model you chose, such as Claude on your plan, your own API key, a local model server, or CMEM Pro on cmem.ai.
The condensed observations land in ~/.claude-mem, and the next session starts with the recent ones in its context.

Wordcell

You or your agent write a note into the vault, a folder of Markdown on your machine.
You commit it, and the note travels wherever the repository goes.
In the next session, the agent’s client starts wordcell mcp on your machine and the agent searches the files. Exact search uses no model. The default hybrid search also runs a local embedding model, downloaded on first use, and hosted reranking runs only when you ask for it.

Who should pick which

The deciding question is whether you want memory to build itself, or memory that someone writes and reviews.

Choose Claude-Mem when

You want memory to build up without asking the agent to save anything.
You work in one repository on one machine and want recent work loaded into each new session.
You accept a model reading every session, on your Claude plan, your own key, a local model server, or its paid service.
You want the same automatic capture in several agents it has installers for, such as Claude Code, Codex, and Cursor.

Choose Wordcell when

You want memory you can read, edit, and review in Git, like code.
You want the memory to come with every clone of the repository, on any machine.
You want saving and exact search to work without a model or an account.
You want notes tied to code paths, so an agent changing a file gets the notes for it.

Run both

The two tools write to different places, so they can share one agent.

Claude-Mem fills its own database, and Wordcell reads only the vault you name with --root.
Claude-Mem’s capture hook matches every tool, so with both installed it also records Wordcell’s tool calls and their results. Its CLAUDE_MEM_SKIP_TOOLS setting lists tools to leave out.
Keep the decisions you want in the repository as Wordcell notes, and let Claude-Mem hold the session history.

Connect Wordcell to your coding agent · Set up a vault · Compare more tools