Projectmem: one MCP server for every project, and a release with no push behind it
Open-source coding agents memory. Records issues, attempts, fixes and decisions, then warns your agent before it repeats an approach that already failed. Native MCP server for Claude Code, Cursor, Antigravity and Codex. 100% local, no cloud, no telemetry. MIT.
At a glance
- What is it?
- A local-first memory layer for coding agents. It stores typed events in a plain directory inside your repository and warns you before you commit an approach that already failed. Three runtime dependencies, seventeen MCP tools, and one release that landed after the default branch's last commit.
- Who is it for?
- Projectmem is one of the more honest agent memory tools to read, because it says what it stores, where it stores it, and what it refuses to send anywhere, and its install path is two commands. Before adopting it, check which version you actually get, since the newest release and the branch tip do not line up, read the pre-commit warning on a small repository first, and treat the token-saving and experience claims as unmeasured until you compare them against your own sessions.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 14 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- riponcm/projectmem882 ★ · +82
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| riponcm/projectmem | 800 | 882 |
82 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
One server, no root, and a cold start requirement
Since version 0.3.0 a single MCP server serves every project, which is what makes the upgrade from 0.1.x or 0.2.x a one-time step. The configuration block is four fields:
"mcpServers": {
"projectmem": {
"command": "/absolute/path/to/python",
"args": ["-m", "projectmem.mcp_server"]
}
}The absence of a --root or a cwd argument is called out as the reason it can serve everything: the server discovers the repositories itself rather than being pointed at one. The command is an absolute Python path, and pjm init prints the block with your own path filled in.
Two operational notes come with it. The client has to be fully restarted, because MCP servers only load on a cold start, and pjm doctor will flag any client still pinned to a single repository, which the documentation names as the most common reason a newly registered project stays invisible to an agent.
Five clients are named for this setup: Claude Desktop, Claude Code, Cursor, Antigravity and Codex.
A doctor command that walks your home directory
The setup flow asks you to find the projects you already have before registering them, and pjm doctor does the searching. It looks where code lives, naming ~/Developer, ~/code, ~/projects, your cloud folders, and every drive on Windows, then lists projects that already have memory but are not registered yet.
Anything it misses can be added by hand:
pjm project register "/Users/you/Developer/repos/ossdrop"and the bulk action is pjm doctor --fix, which registers what it found. Running doctor again after editing a client config is how you confirm the change landed.
The network story is handled in the same place. projectmem makes no network calls by default, and the only one available is an update check you turn on yourself with pjm doctor --online, or once a day with --auto. That matches the wider claim: memory lives in a plain .projectmem directory in your repository, with no cloud, no account and no telemetry.
Typed events instead of chat history, and the warning they enable
The distinguishing design choice is the record format. projectmem stores typed events, and the five types are issues, attempts, fixes, decisions and notes. The documentation draws the contrast with chat-history memory tools directly.
That typing is what makes the signature feature possible: a pre-commit warning that fires before you repeat an approach that already failed in this project. An agent that only has a transcript would have to infer the failure; an agent that has an attempt record and a fix record can be stopped at the commit boundary.
The storage is a plain file. Events land in .projectmem/events.jsonl, which is also the input to pjm visualize, the command that reports every case the project solved, what failed on the way, and what it saved, generated locally from that file.
The tutorial is explicit that seeing the warning fire is part of the walkthrough, which is the right thing to test before trusting it.
v0.3.4 shipped, and the branch tip is still v0.3.3
Three releases are visible, and each carries a sentence about what went wrong before it. v0.3.2 is titled Windows watch support, and a doctor that tells the truth. v0.3.3 is titled precheck stops getting slower, and Windows works end to end. v0.3.4 is titled memory that stays small, and warnings you can trust.
The timestamps are the interesting part. v0.3.3 was published on 2026-09-15 at 00:15:20 UTC, about half a minute after the last push to the default branch at 00:14:50. v0.3.4 was published on 2026-10-02 at 23:48 UTC, and the default branch has not been pushed to since that same September moment, even though the project metadata reads version 0.3.4.
So the newest release and the newest commit on main are not the same thing, and the release notes for 0.3.4 describe a time-zone regression inherited from 0.3.3 alongside fixes for three community reports.
The staleness rule is the concrete casualty. A warning whose freshness depended on your time zone is a warning that could be wrong by hours depending on where the machine was set.
Three runtime dependencies and twenty-seven keywords
The build is hatchling, and the dependency list is short enough to read in one breath: typer for the command line, the mcp package bounded at 1.0 to 3, and watchdog for filesystem events. Development extras add pytest and ruff.
Three dependencies is a real argument for the local-first claim, since nothing in that list can phone home. watchdog is also the reason the pre-commit warning can exist at all, because it is what notices a file changing in the first place.
The keyword list is the unusual part, and it is annotated. Twenty-seven keywords run from coding-agent-memory and agent-memory through mcp, mcp-server, claude, cursor and codex, with a comment above the block explaining that single generic words such as ai and memory cannot rank on PyPI, that coding-agent-memory is the canonical phrase the documentation and site use, and that it contains agent-memory as a substring so it carries the category.
The classifiers list Python 3.10 through 3.14 and mark the project as Development Status 5, Production/Stable, which is a confident label for a 0.3 release.
A paper, a citation file and an llms.txt, with the guides on a website
The repository root is short: .github/, .gitignore, CHANGELOG.md, CITATION.cff, LICENSE, README.md, SECURITY.md, TUTORIAL.md, brand/, llms.txt, pyproject.toml, src/ and tests/.
Two of those entries are worth reading for what they signal. CITATION.cff means the project can be cited through the citation standard rather than by hand, and llms.txt is the machine-readable index intended for language models, which is an unusual file to find at the root of a Python package.
The documentation is split by audience. The README carries the quick start and the concept explanation, TUTORIAL.md is a fifteen minute walkthrough that includes the pre-commit warning firing, CHANGELOG.md holds the release history, and the complete setup guide with per-client MCP instructions lives on the website rather than in the repository.
The research side is one paper, arXiv 2606.12329, titled as a local-first event-sourced memory and judgment layer for AI coding agents, covering design, a Memory-as-Governance framing, a capability comparison and a 207-event dogfooding study.
The pitch is experience, and the measurement is elsewhere
The positioning is a single contrast: we do not make AI smarter, we make it experienced. The problem section argues that the model is not the problem, that the architecture is, and that stateless models need a memory cortex. The named cost is that agents forget yesterday's decisions, repeat failed debugging attempts, and burn millions of tokens reconstructing context from raw source files.
None of those numbers are measured in the README. There is no token count for a before and after, no session comparison, and no figure attached to the claim about millions of tokens. The one study named anywhere is the 207-event dogfooding run inside the arXiv paper, which is where a reader has to go for evidence.
The Solution section is also where the text stops, mid-sentence, after Git tracks what changed and projectmem tracks why it changed, what.
What the README does support without measurement is structural: the file layout, the tool count, the dependency list, and the rule that nothing leaves the machine unless you ask it to.
Editorial conclusion
Projectmem is one of the more honest agent memory tools to read, because it says what it stores, where it stores it, and what it refuses to send anywhere, and its install path is two commands. Before adopting it, check which version you actually get, since the newest release and the branch tip do not line up, read the pre-commit warning on a small repository first, and treat the token-saving and experience claims as unmeasured until you compare them against your own sessions.
Frequently asked questions
What does projectmem store inside a project?
Memory lives in a plain .projectmem directory inside your repository as typed events rather than chat history: issues, attempts, fixes, decisions and notes. Events are written to .projectmem/events.jsonl, and pjm visualize generates a local view of what the project solved and what failed along the way.
How many MCP tools does projectmem expose, and to which clients?
Seventeen tools through a native MCP server, available to Claude Code, Claude Desktop, Cursor, Antigravity and Codex. Since version 0.3.0 one server serves every project, configured with no --root and no cwd, and pjm doctor flags a client still pinned to a single repository.
Does projectmem make any network calls?
Not unless you ask. There is no cloud, no account and no telemetry, and the only network call available is an update check you enable yourself with pjm doctor --online, or turn into a once-a-day check with --auto.
What is the pre-commit warning in projectmem?
It is the signature feature: because projectmem stores typed events rather than transcripts, it can warn you before you commit an approach that already failed in that project. The fifteen minute TUTORIAL.md walkthrough includes seeing it fire.
What does the pjm doctor command check?
It searches where code lives, including ~/Developer, ~/code, ~/projects, your cloud folders and every drive on Windows, and lists projects that have memory but are not registered yet. pjm doctor --fix registers them, pjm project register adds one by hand, and re-running doctor confirms the client configuration.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/riponcm-projectmem)Star history chart for your README
[](https://hysenlabs.com/projects/riponcm-projectmem)Was this useful?