Projectmind

projectmind

Persistent, compact project memory for AI coding agents. Your agent reads one small digest instead of re-scanning the codebase every session — and a local ledger shows you exactly how many tokens (and dollars) that saved.

[](https://github.com/Nodemint-dev/projectmind/actions/workflows/ci.yml)

-success)

Every AI coding session starts the same way: the agent has no memory of your project, so it re-reads files, re-derives your architecture, or asks you to re-explain decisions you made months ago. You pay for that in tokens, time, and wrong guesses — every single session, on every machine, for every teammate.

projectmind fixes this with a structured, persistent project map (.projectmind/map.json): your modules, their dependencies, your architectural decisions, your conventions, your domain glossary. The agent reads a ~400-token digest first, drills into single nodes only when needed, and writes back what it learns — so the next session (yours, a teammate's, or a different AI tool's) starts already knowing the project.

The digest scales with your project's conceptual size (modules, decisions), not its byte size — so the savings grow with the repo.

Quick start — one command

npm install -g @nodemint/projectmind
cd your-repo
projectmind init

That single init does everything: scaffolds the map, seeds it from your repo layout, detects which AI agents are installed on your machine and wires each one (MCP server + a rules file carrying the live map), adds .projectmind/ to your .gitignore (the map stays local — nothing to push), and installs the git hook that keeps the map fresh on every commit.

Then the one step no tool can do for you: restart your agent (or start a new chat session). AI agents load their config and rules files at session start — a chat that was already open keeps its old context, and every session after that has the map automatically. Using the VS Code savings extension? Cmd+Shift+P → "Developer: Reload Window" once.

Want it available in every future project without even running init?

projectmind setup --global        # registers the MCP server once, user-wide

(projectmind init --bare scaffolds only, and projectmind setup re-wires agents on demand, if you prefer the pieces separately.)

See what it saves you — measured, not promised

The MCP server keeps a local savings ledger: every time your agent reads the map instead of scanning files, it records the tokens actually served vs. the estimated tokens of the files the agent would have read instead.

Check it anytime:

$ projectmind savings
projectmind savings (estimated — local ledger, never leaves this machine)

  Total saved:   ~20.5k tokens across 22 map read(s)
  ≈ $0.0614 at sonnet-tier input pricing ($3/MTok, as of 2026-06)
  Today:         ~20.5k tokens (22 read(s))

  By tool:
    mind_digest      12 reads   ~18.5k saved
    mind_query        6 reads   ~1210 saved
    mind_context      4 reads   ~712 saved

  Methodology: tokens ≈ ceil(bytes/4); baseline = files the agent would have read instead.

The dollar figure needs zero configuration: it defaults to Sonnet-tier input pricing from a built-in table of published rates (Haiku $1 / Sonnet $3 / Opus $5 / Fable $10 per MTok input, as of June 2026), and says exactly which assumption it used. If you run a different model, set savings.model (or an exact savings.inputPricePerMTok) in .projectmind/config.json — but nobody has to.

(Real output from a scripted day of agent work on the sample project in test/fixtures/ — 12 session starts, 6 module drill-ins, 4 task-context reads. Reproduce the per-session number with npm run benchmark.)

VS Code status bar: the [integrations/vscode](integrations/vscode) extension shows a live ✦ ~20.5k tokens saved counter that reads the same local ledger. Zero dependencies, zero network.

The ledger is honest by design: every number is labelled an estimate (ceil(bytes ÷ 4), the rough English+code average), savings are floored at zero, and the file is gitignored — it's your private data, on your machine, deletable at any time.

One deliberate gap, stated plainly: the ledger only counts tool calls (mind_digest, mind_query, mind_context). The embedded digest described below never requires a tool call — that's the whole point — so its savings don't show up as a number here. projectmind savings names which rules files carry it instead of inventing a token count we can't actually observe.

Session handoff — pick up exactly where you left off

The thing every agent session loses is working state: what you were in the middle of, what's next, the gotcha you just discovered. Code graphs can't capture it; chat history dies with the session. projectmind carries it over:

$ projectmind digest
# shopflow-api — project map
Order-management HTTP API with JWT auth, backed by PostgreSQL.
Stack: node, express, postgres

## ⏪ Handoff from last session (2026-07-02)
Adding refund support to orders-route; next: write the refund tests

Your agent calls mind_handoff before the session ends (or before its context gets compacted); the note leads the very next digest, so the next session — tonight on your laptop, tomorrow on your desktop, or a different AI tool entirely — resumes in one read instead of re-deriving the task. Notes live in your gitignored local overlay: personal working state, never committed, cleared with mind_handoff({clear: true}) when done. Humans can use it too: projectmind handoff "note".

Works everywhere your team works

No language assumptions, no platform assumptions: the full test suite (85 tests, including the git-hook end-to-end and offline-guarantee tests) runs in CI on Linux, macOS, and Windows × Node 18/20/22. Paths, globs, atomic renames, and the installed git hook are exercised on all three. The map format is plain JSON — nothing OS-specific is ever written to your repo.

Verified against real open-source repos (fresh clone → projectmind init --seed, 2026-07-02):

| Repo | Stack detected | Seeded nodes | Repo size (est.) | Digest | |------|---------------|--------------|------------------|--------| | expressjs/express | node | examples, lib, test | ~175k tokens / 152 files | ~86 tokens | | pallets/flask | python | examples, src, tests | ~153k tokens / 104 files | ~111 tokens | | flutter/pinball | dart, flutter | lib, packages, test, web | ~513k tokens / 637 files | ~99 tokens |

That's the core scaling property in the wild: the digest tracks a project's conceptual size, staying ~100 tokens whether the repo is 150k or 500k tokens. (A freshly seeded digest is a starter skeleton — a curated map with decisions and conventions lands around 400 tokens, like the benchmark fixture. We quote the honest per-session savings number — 78.9% — from the benchmark, not from these whole-repo ratios.)

How it works

flowchart LR
    subgraph agents["Your AI agents"]
        A1[Claude Code]
        A2[Cursor]
        A3[Gemini / Codex / Copilot]
    end
    subgraph pm["projectmind"]
        MCP["MCP server\n(stdio, 6 tools)"]
        CORE["core\n(atomic writes, schema-validated)"]
        HOOK["git post-commit hook\n+ watch mode"]
    end
    subgraph repo["Your repo"]
        MAP[".projectmind/map.json\nnodes · edges · decisions\nconventions · glossary"]
        DIG[".projectmind/digest.md\n+ embedded in CLAUDE.md etc."]
    end
    LEDGER[".projectmind/ledger.json\nlocal savings ledger (gitignored)"]

    A1  MCP
    A2  MCP
    A3  MCP
    MCP --> CORE
    HOOK --> CORE
    CORE  MAP
    CORE --> DIG
    MCP --> LEDGER

Two update paths keep the map current without burning tokens:

• Deterministic: the git hook (and optional projectmind watch) map changed

files to modules via globs and bump freshness. Pure local computation, zero LLM.

• Agent-recorded: when the agent makes an architectural decision or learns a

convention, it calls mind_update — a few tokens once, instead of re-discovery every session.

There is deliberately no background LLM summarization — that would burn the tokens this tool exists to save.

What the agent actually reads

# shopflow-api — project map
Order-management HTTP API with JWT auth, backed by PostgreSQL.
Stack: node, express, postgres

## Active
- **auth**: Issues/verifies JWT access tokens; exposes requireAuth Express middleware. [active]
- **orders-route**: Create and list orders for the authenticated user; totals in integer cents. [active]

## Modules
- **db**: PostgreSQL pool wrapper: query() helper and withTransaction().
…

## Key decisions
- Store money as integer cents, never floats. (2026-06-10)
- JWT for stateless auth instead of server sessions. (2026-06-15)

## Conventions
- Never read process.env outside src/config.js.
…

~412 tokens for the whole thing. The long fields (notes, rationale, file lists) are excluded from the digest and only surface through mind_query — cheap by default, detail on demand.

MCP tools

| Tool | Purpose | |------|---------| | mind_digest() | Read first every session. The compact map. | | mind_context({files}) | Task-scoped. Only the modules a task touches + their direct deps. The cheapest read when you know the files. | | mind_query(node) | Full detail on one node: files, notes, edges. | | mind_search(term) | Find nodes / decisions / glossary by keyword. | | mind_update(delta) | Record a structural change, decision, or convention. | | mind_handoff(note) | Leave a "resume here" note that leads the next session's digest. | | mind_stats() | Map size, digest cost, and your savings ledger. |

One-command agent wiring — and it doesn't depend on the model choosing to read it

projectmind setup idempotently writes the MCP config and a workflow rules block for each agent — merging into existing configs, never clobbering (an unparseable config is backed up and skipped).

Here's the part that matters: MCP tools are a nudge — no server can force a model to call one, so an agent can (and sometimes will) skip mind_digest on a plain "explain this project" question, especially mid-task. But rules files (CLAUDE.md, .cursorrules, .windsurfrules, GEMINI.md, AGENTS.md, .github/copilot-instructions.md) are loaded into every agent's context unconditionally, with zero model choice involved — that's the actual mechanism, not a coincidence. So setup doesn't just write instructions to call mind_digest; it embeds the live digest itself between ` markers, and every subsequent mind_update, CLI edit, git-commit hook run, or watch` save re-syncs it automatically. The project map is present in context from message one, on every agent, with no tool call required at all — the same reliability class as a context-injecting hook, without needing one.

(Only the repo-committed map is ever embedded — never your gitignored local overlay or handoff notes, so nothing personal leaks into a file you commit.)

| Agent | MCP config | Rules file | |-------|-----------|------------| | Claude Code | .mcp.json | CLAUDE.md | | Cursor | .cursor/mcp.json | .cursorrules | | Windsurf | .windsurf/mcp.json | .windsurfrules | | Gemini CLI | .gemini/settings.json | GEMINI.md | | Codex / generic | — | AGENTS.md | | GitHub Copilot | — | .github/copilot-instructions.md |

Target one with projectmind setup --agent cursor. Manual wiring is one line everywhere:

{ "mcpServers": { "projectmind": { "command": "npx", "args": ["-y", "@nodemint/projectmind", "mcp"] } } }

Install once, use everywhere — global setup

projectmind setup above is per-project. Run projectmind setup --global instead to register the MCP server once, so it's available in every project you open from then on — no per-repo .mcp.json, no re-running setup. Supported: Claude Code, Cursor, Windsurf, Gemini CLI.

projectmind setup --global

For Claude Code this shells out to its own claude mcp add --scope user (the same mechanism tools like codegraph use for global registration) rather than hand-editing its internal config file. For the others it merges into their global config (~/.cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, ~/.gemini/settings.json) the same idempotent way the per-project setup does.

Two things this can't do, by design:

• Rules files stay per-project. CLAUDE.md / .cursorrules / etc. are

checked-in project conventions — there's no sane "global" version of them, so --global only registers the MCP server, not the nudge-text. Without a project's .projectmind/ map, mind_digest just returns an empty-but-valid digest — harmless, but the agent won't be as strongly reminded to call it.

• A chat session already open won't see it. Every MCP client (Claude Code,

Cursor, etc.) loads its server list once at session start — that's an MCP client behavior, not something any server can override. Global scope means every new session in every future project has it automatically; it does not retroactively add it to a conversation already in progress. Restart the agent once after running --global, and you're done for good.

Where projectmind fits (and what it deliberately isn't)

| | Structural code graphs (e.g. codegraph) | Coding-policy plugins (e.g. ponytail) | projectmind | |---|---|---|---| | Captures | symbols, call paths — the how | how agents should write code | intent, decisions, conventions — the why | | Source | parsed from code | prompt policy | curated by you + your agent | | Can it know why you chose JWT over sessions? | no | no | yes | | Carries your working state across sessions | no | no | yes (handoff) | | Shows you what it saved | no | no | yes (local ledger) | | Works offline | varies | yes | always (CI-enforced) |

You can't parse "money is always integer cents" or "we chose JWT for horizontal scaling" out of source code. That's the layer projectmind owns.

So it complements structural tools instead of cloning them: if a .codegraph/ index exists in your repo, mind_digest, mind_query, and mind_context automatically point the agent to codegraph for symbol-level detail. No dependency, no duplication — projectmind is the why, codegraph is the how.

Keeping the map honest

$ projectmind doctor
Nodes pointing at files that no longer exist:
  - legacy-auth (globs: src/old-auth/**)
Nodes untouched for more than 90 days:
  - reports (last touched 2026-03-12, 112 days ago)

• projectmind doctor — drift detection: dangling file globs, stale nodes.
• projectmind validate — schema integrity + drift warnings.
• projectmind watch — live freshness on save (fs.watch, debounced, local).
• Corrupt map.json? It's backed up to map.json.corrupt- and the session

continues with an empty valid map — the agent never crashes on a bad file.

• Every write is schema-validated and atomic (temp file + fsync + rename), so

a racing agent and git hook can't corrupt the map.

The benchmark (reproduce it yourself)

$ npm run benchmark
Project: sample-project
Baseline (files an agent would read):  ~1953 tokens  (8 files)
projectmind digest:                    ~412 tokens
Savings:                               78.9%  (~1541 tokens/session)

Methodology, stated plainly: the baseline is the concatenated content of the files an agent would plausibly read to orient itself (README + package.json + all src/ files of the committed fixture project); tokens are estimated at ceil(chars ÷ 4). These are estimates, not exact counts — the fixture and the script are in test/, so the number is auditable and reproducible. Savings on a real repo are typically much larger, because real repos have far more than 8 files while the digest stays roughly c

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Nodemint-dev
• Source: Nodemint-dev/projectmind
• License: MIT
• Homepage: https://www.npmjs.com/package/@nodemint/projectmind

Install and usage instructions live in the source repository linked above.