What is MCP Memory Keeper?
MCP Memory Keeper stores what a coding agent is doing as keyed context items (a task, a decision, a progress note) with a category, a priority and a channel, in a local SQLite database at ~/mcp-data/memory-keeper/context.db. Named checkpoints snapshot those items, cached file contents and git status, and context_restore_checkpoint brings them back after Claude Code compacts or restarts. Channels are derived from the git branch by default, so feature/auth-system becomes a channel of at most 20 characters.
The server exposes 38 tools over stdio by default, from context_save, context_get and context_search to session branching, journal entries, a knowledge graph and a context_git_commit tool. A TOOL_PROFILE variable cuts that to 8 (minimal) or 22 (standard). Installation is claude mcp add memory-keeper npx mcp-memory-keeper; no account or key is involved.
The repo was created in June 2025 by Mark Kreyman, is MIT-licensed and has about 136 stars, with almost all of its commits from one maintainer. The latest npm version is 0.14.1 from 2 September 2026 (GitHub releases stop at v0.12.1), and the package logged 140 downloads in the week to 22 October. It is not listed in the official MCP registry.
How to install MCP Memory Keeper
claude mcp add memory-keeper npx mcp-memory-keeperIs MCP Memory Keeper worth it?
Memory Keeper is built around a real problem with long agent sessions: when the context window compacts, decisions and half-finished plans vanish. Its answer is deliberate rather than automatic. The README's own workflow has you ask Claude to save a checkpoint when the conversation gets long and to restore it by key after the reset; nothing hooks into compaction for you, and recall depends on instructions in CLAUDE.md or a custom command.
The tool list is the main cost. The default full profile lists 38 tools, each with a schema the model reads on every turn, for a job the reference memory server covers with nine. The maintainer acknowledges this with profiles: minimal keeps 8 (sessions, save, get, search, status and the two checkpoint tools), and that is where most people should start. Version 0.12.2 also fixed the full profile breaking OpenAI-compatible providers, a sign of how far the surface has spread.
Some features are lighter than their names. context_semantic_search does not use a language model: in the source it hashes character trigrams and words into a 384-number vector and compares those, so it matches similar spelling rather than meaning. The knowledge-graph and multi-agent tools run locally over the same saved text. Storage is sounder: SQLite in WAL mode, an absolute database path since 0.12.x, and import and export confined to one directory.
Two security points deserve attention. context_git_commit runs git add . and commits everything in the project directory, so an agent can stage files you did not mean to commit; leave it out of your profile unless you want that. And version 0.13.0 (June 2026) fixed context_import reading any file the server could reach, a hole a prompt-injected agent could have used to pull secrets into memory. The fix is in, but an open issue still asks the maintainer to enable private vulnerability reporting.
Against the reference memory server, the trade is structure for weight. Memory gives you one readable JSONL file and an entity graph; Memory Keeper gives categories, priorities, branch-aware channels and checkpoints in a database you query with sqlite3. With about 140 npm downloads a week and one maintainer, it is a small project, actively released through September 2026.
Where it earns its place
A solo developer doing a multi-day refactor in Claude Code
Pick up where the agent left off after compaction or a restart
Run with TOOL_PROFILE=minimal, add a CLAUDE.md rule to save decisions and progress, and ask for context_checkpoint before the window fills. After the reset, context_restore_checkpoint returns the saved items. It is free; the cost is remembering to ask.
A developer running two Claude Code sessions in parallel, one researching and one implementing
Let one session hand findings to the other
Both point at the same DATA_DIR; items are visible across sessions by default, and a shared channel keeps the hand-off on one topic. There is no access control between sessions, so anything one saves the other can read.
An engineering lead approving MCP servers for a 20-person team
Decide whether staff may install it on work laptops
Data stays in a local SQLite file and the source makes no network calls of its own, so there is no processor to assess. The checks are the git commit tool (exclude it with a custom profile in ~/.mcp-memory-keeper/config.json), versions before 0.13.0 with the import file-read bug, and a one-maintainer project with no private disclosure channel.
Before you commit
- Setup
- claude mcp add memory-keeper npx mcp-memory-keeper, or the npx snippet in the Claude Desktop config; the README says under 30 seconds. Data goes to ~/mcp-data/memory-keeper/ unless DATA_DIR is set. Users upgrading from before 0.12.x must move context.db from their old working directory.
- Integrations
- 38 tools by default, all named context_*: context_session_start, context_save, context_get, context_search, context_checkpoint, context_restore_checkpoint, context_prepare_compaction, context_export, context_import and more. Most write to the database; context_git_commit writes to your repository. The README's examples prefix them with mcp_, which is not the registered name.
- Data and privacy
- Everything is in one SQLite database plus an exports directory. Saved items are shared across sessions by default. Anything the agent reads, including injected text from a web page or file, can be saved and replayed in later sessions. Version 0.13.0 fixed an arbitrary local file read through context_import (#35).
- Cost at scale
- Free, with no rate limits. The real cost is context: the full tool list on every turn, and responses capped at 25,000 tokens by default (MCP_MAX_TOKENS), which a large context_get can fill.
- Limits
- stdio only, one user. No automatic checkpoint at compaction; the agent must be told. Search is text and trigram matching. The better-sqlite3 native module failed to load after Node upgrades until 0.14.1, and the README lists retention policies and multi-user support as future work.
What's good
- Checkpoints capture items, cached files and git status together
- Channels follow the git branch automatically
- Tool profiles cut 38 tools to 8 or 22
- Local SQLite; no account, key or network service
What's not
- 38 tools by default inflate every prompt
- context_git_commit stages and commits the whole project directory
- "Semantic" search is hashed trigrams, not embeddings
- Single maintainer; about 140 npm downloads a week
Best for
Claude Code users who run long sessions and are willing to tell the agent when to checkpoint and restore.
MCP Memory Keeper alternatives
MCP Memory Keeper: common questions
- Is MCP Memory Keeper worth it?
- A thoughtful answer to losing work at compaction, wrapped in 38 tools and a commit command most users should switch off. We rate it 3.4 out of 5.
- How much does MCP Memory Keeper cost?
- Free and MIT-licensed.
- What are the downsides of MCP Memory Keeper?
- 38 tools by default inflate every prompt; context_git_commit stages and commits the whole project directory; "Semantic" search is hashed trigrams, not embeddings; single maintainer; about 140 npm downloads a week.
- Who is MCP Memory Keeper best for?
- Claude Code users who run long sessions and are willing to tell the agent when to checkpoint and restore.
- How much does MCP Memory Keeper cost for a team or in production?
- Free, with no rate limits. The real cost is context: the full tool list on every turn, and responses capped at 25,000 tokens by default (MCP_MAX_TOKENS), which a large context_get can fill.
- Is MCP Memory Keeper safe for company data?
- Everything is in one SQLite database plus an exports directory. Saved items are shared across sessions by default. Anything the agent reads, including injected text from a web page or file, can be saved and replayed in later sessions. Version 0.13.0 fixed an arbitrary local file read through context_import (#35).
- What are the best MCP Memory Keeper alternatives?
- The closest alternatives we have reviewed are Memory (reference server) and Project RAG.
Answers checked Oct 26, 2026.