@nodemint/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.

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, git-committed 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

npm install -g @nodemint/projectmind        # installs the `projectmind` CLI
cd your-repo
projectmind init --seed           # scaffold + propose a starter map from your repo layout
projectmind setup                 # wire the MCP server + rules into every agent you use
projectmind install-hook          # optional: auto-update module freshness on commit

That's it. Your agent now calls mind_digest at the start of a task instead of scanning files.

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 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.

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 (63 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 (git-committed)"]
        MAP[".projectmind/map.json\nnodes · edges · decisions\nconventions · glossary"]
        DIG[".projectmind/digest.md\nreviewable in PRs"]
    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

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

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-<ts> 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 constant-size.

CLI reference

projectmind init [--seed]                       # scaffold .projectmind/ (+ starter map from repo layout)
projectmind seed                                # propose starter nodes (never overwrites curated ones)
projectmind setup [--agent <name>]              # wire MCP + rules into agents (default: all)
projectmind digest                              # print the digest (what the agent sees)
projectmind context [--files a,b] [--node id] [--term t] [--depth 1]
projectmind query <id>                          # full node detail
projectmind search <term>
projectmind add-node <id> "summary" [status]    # active | stable | deprecated
projectmind add-edge <from> <to> <rel>
projectmind decide "text" ["rationale"]
projectmind convention "text"
projectmind handoff ["note"] [--clear]          # leave/show/clear the resume-here note
projectmind stats                               # sizes + estimated digest tokens
projectmind savings                             # your local savings ledger
projectmind validate                            # map integrity + drift warnings
projectmind doctor [--stale <days>]             # drift report
projectmind watch                               # live freshness updates on save
projectmind install-hook                        # git post-commit auto-updater

Options: --local (per-developer overlay), --root <dir>

The map, in 20 seconds

{
  "version": 1,
  "project": { "name": "shopflow-api", "stack": ["node", "express", "postgres"] },
  "nodes": {
    "auth": {
      "type": "service",                      // module | component | service | doc | concept
      "summary": "Issues/verifies JWTs; requireAuth middleware.",  // one line — the token budget
      "files": ["src/auth.js"],               // globs; the hook maps commits → nodes with these
      "status": "active",                     // active | stable | deprecated
      "notes": "Longer detail — never in the digest, only via mind_query."
    }
  },
  "edges": [{ "from": "auth", "to": "config", "rel": "depends-on" }],
  "decisions": [{ "id": "d3", "text": "JWT over sessions.", "rationale": "…", "date": "2026-06-15" }],
  "conventions": ["Money is integer cents."],
  "glossary": { "SKU": "Stock Keeping Unit." }
}

Two scopes: the committed map.json (shared with your team) and an optional gitignored map.local.json overlay for personal context ("mid-refactor on X"). digest.md is committed too, so map changes show up as readable diffs in PRs.

Trust FAQ

Does it send my code anywhere? No — and you don't have to take our word for it. A CI test (test/offline.test.js) scans the entire source tree and fails if any network API (http, net, fetch, sockets, …) appears. Two runtime dependencies (@modelcontextprotocol/sdk, picomatch), no telemetry, no LLM calls, no accounts.

What gets committed? map.json, digest.md, config.json. Your personal overlay (map.local.json) and your savings ledger (ledger.json) are gitignored automatically.

Are the savings numbers real? They're honest estimates, clearly labelled, with the methodology printed next to every number and a reproducible benchmark in the repo. We'd rather under-claim than exaggerate.

Can the map rot? The git hook and watch mode keep file↔module freshness current for free; doctor flags dangling and stale nodes; validate checks integrity. And because the digest is committed, drift is visible in code review.

Does it lock me into one AI tool? No. Any MCP-capable agent can use it, the rules files cover the rest, and the map itself is plain JSON any tool can read.

Development

npm install
npm test              # 63 tests: schema, atomicity, corruption self-heal, scopes,
                      # globs, MCP round-trip, ledger, handoff, offline guarantee
npm run benchmark     # prints the estimated savings number

MIT © contributors. See CHANGELOG.md.