v0.1.5 — Reliability Release

Same features, fewer ways to break. Three community-contributed fixes make projectmem safer to run everywhere — especially on Windows. No new dependencies, no schema changes.

Fixes

• pjm fix --issue / record_fix(issue_id=…) — close the issue you mean. Previously fix always closed the "current" issue, so fixing an older one could silently close the wrong issue. You can now target a specific issue by ID (pjm fix --issue 0001; IDs normalize 1/001/0001). Closes #3. (#2)
• MCP git helpers no longer hang stdio sessions — git subprocesses now run with detached stdin (DEVNULL) and bounded timeouts, fixing a class of Windows MCP hangs. (#4)
• Console output is encoding-safe — pjm brief and pjm precheck no longer crash on non-UTF-8 (cp1252) Windows consoles and git hooks. (#5)

🙏 Thanks

All three fixes contributed by @hanley-development — thank you!

Install

```
pip install -U projectmem
```

Full Changelog: v0.1.4...v0.1.5

v0.1.4 — The Accountable-Judgment Release

The accountable-judgment release — memory that flags its own staleness instead of silently trusting (or deleting) it. Six features, ~330 lines, no new dependencies, the same 14-tool MCP surface.

📦 pip install -U projectmem · 📰 Release post · 🎓 Research paper (arXiv:2606.12329) · 🎬 Tutorial video

✨ New

• Stale-memory detection — flag, never delete. Every decision/fix/note that cites a file is cross-checked against that file's git history. When the file moved on (3+ commits, or it no longer exists), pjm precheck and the MCP precheck_file tool flag the memory — "predates 7 commits to auth.py — confirm or supersede." Nothing is deleted; a human decides. Deterministic git log counts, no embeddings, no daemon.
• Superseded-decision marking. pjm decision "switch to argon2" --supersedes <event-id> (also on MCP add_decision) retires an old decision without rewriting history: it leaves summary.md, stays in the append-only log, and shows in pjm search tagged (superseded). Search now prints event ids.
• pjm brief — one-screen session-start briefing: active warnings, possibly-stale memories, open issues, recent decisions, stack gotchas, and your prevention score with a weekly delta.
• Precheck snooze. pjm precheck --snooze 2h pauses warnings for a bounded window instead of pushing you to --no-verify. The snooze is itself logged — even the silence is audited. --unsnooze restores early.
• Failed approaches listed in the warning. Precheck now shows the dead ends themselves ("✗ tried CSS contain:layout"), and pjm search --failed-only lists a project's full catalogue of dead ends.
• pjm export --claude-md (and --cursor) — compiles live decisions (stale ones flagged), gotchas, and a "Do NOT retry — these already failed" list into a marked block in CLAUDE.md / .cursorrules. Agents without MCP inherit the project's judgment.
• Dashboard Overview tab — pjm visualize opens on an all-at-a-glance view (failure heatmap, A+→F prevention gauge, project graph with failure rings, typed-event timeline), re-themed light.

🐛 Fixed

• Walk-up discovery no longer mistakes ~/.projectmem (the machine-wide global store) for a project — found by dogfooding 0.1.4 on this very repo.
• pjm precheck payment.py auth.py — named files now work directly from the CLI.
• pjm search payment.py matches the location field, so per-file lookups behave like precheck.

📋 Compatibility

One new optional event field (supersedes); existing logs parse unchanged; MCP surface unchanged (14 tools). After upgrading run pjm hooks install once per project. 23 new tests — 81 passing.

Full changelog: https://projectmem.dev/changelog

v0.1.3 — quality + privacy + conda/venv hook fix + init UX

Six focused improvements:

• Schema enrichment — all 14 MCP tools now have parameter descriptions + schema constraints. Projected Glama score 75% → ~90%.
• Secret redaction on write — API keys, tokens, PEM blocks scrubbed from event text before they touch disk. Default-on, PROJECTMEM_NO_REDACT=1 to skip.
• L-047: conda/venv hook fix — pre-commit warning now works under conda/pyenv/venv. The killer feature had been silently no-op for most Python users.
• L-048: stack auto-detect — pjm init pre-populates PROJECT_MAP.md from your manifests (pyproject / package.json / Cargo / go.mod). Setup Mode becomes refinement, not greenfield.
• L-049: MCP config printed at init — copy-pasteable JSON block with paths for Claude Desktop / Cursor / Antigravity / Codex.
• L-050: silent post-commit auto-capture — no more output printing over your shell prompt after git commit.

Plus TUTORIAL.md — 15-minute step-by-step walkthrough linked from the README.

Install: pip install -U projectmem
Tutorial: TUTORIAL.md
Full changelog: CHANGELOG.md

v0.1.1 — First Stable Release · Cross-Project Memory Verified

projectmem v0.1.1 — First Stable Release

The local-first memory layer for AI coding agents. projectmem captures what your AI learned (and what didn't work) so the next session — and the next project — starts experienced instead of from zero. 100% local. No cloud. No telemetry. No account.

After months of dogfooding, 46 documented lessons, a 22-item polish-pass, four MCP clients verified end-to-end, and three language ecosystems put through a real cross-project memory test cycle — v0.1.1 is the version we're comfortable putting on PyPI as a stable release.

pip install projectmem
cd your-project
pjm init

Two commands and your AI is no longer amnesiac.

---

What projectmem does, in one sentence

It gives your AI persistent memory — about your bugs, your decisions, the approaches you've already tried, and the libraries that have bitten you before — accessible through 14 native MCP tools that work with Claude Desktop, Cursor, Antigravity, and Codex out of the box.

What's in this release

The intelligence layer

• Pre-commit warnings — fires before a commit if you're about to repeat a failed approach, touch a high-churn file, or revisit an unresolved issue. The only AI memory tool that prevents mistakes instead of just remembering them.
• Real-time file watcher — auto-starts on pjm init. Detects rapid edits to the same file (debugging sessions) and logs churn events automatically. Battery-aware, gitignore-aware.
• Smart context injection (pjm wrap) — wraps your AI agent with a token-budgeted memory block before the session starts. Inject into CLAUDE.md, .cursorrules, or clipboard.
• Prevention Score (pjm score) — quantifiable A+ → F letter grade backed by debugging hours saved, tokens prevented, and dollars protected. CTO-readable ROI.

Cross-project memory — verified across 3 languages

A library gotcha learned in one project surfaces automatically in any other project on the same stack, with source-project attribution. Three wiring gaps closed in this release:

• Auto-promote wiring restored. Failed attempts and gotcha-prefixed decisions now consistently propagate to the machine-wide store. Wired through storage.append_event so every write surface — MCP, CLI, git hooks — promotes uniformly.
• Language parity. Self-curating library cache at ~/.projectmem/global/.promotable.json means Go, Rust, Java, Ruby, and mobile projects accumulate cross-project knowledge the same way JavaScript and Python ones did. A Go user's gin lessons now propagate exactly like a React user's vite ones.
• Signal filter. Project-setup decisions stay local. Only deliberate lessons — failed attempts, or decisions/notes opening with gotcha: / lesson: / warning: / caution: / pitfall: / avoid: / don't / do not / never / bug: — reach the global store. Signal-to-noise went from 14% to 100% in the real test cycle.

End-to-end verification across proj-react → proj-next (JS inheritance with attribution), proj-python (negative — vite gotcha correctly bounded by stack), and proj-go (gin promotion under the new library cache + signal filter).

Cross-client compatibility

Four MCP clients tested end-to-end. 124 of 124 cells green.

Client	Status	Setup
Antigravity	✅ All checks pass	Native — cwd honored, no workaround
Claude Desktop	✅ All checks pass	Use Auto mode (not Plan); pass root via --root in args
Cursor	✅ All checks pass	Same --root workaround for cwd-ignored bug
Codex	✅ All checks pass	TOML config at ~/.codex/config.toml; medium reasoning required

Each client ships with a documented setup block, captured during real verification.

Operational hardening

• MCP stdio integrity — write tools no longer corrupt the JSON-RPC stream. Five consecutive log_issue / record_attempt / record_fix / add_decision / add_note calls survive cleanly in every client.
• MCP project-root discovery — server walks up the tree to find .projectmem/ (like git does for .git/). Plus --root flag and PROJECTMEM_ROOT env var for explicit pinning.
• No silent issue misattribution — record_attempt after record_fix never latches onto a stale open issue. Marker file + 5-minute time-fence + explicit --issue flag.
• AI workflow alignment — three surfaces (MCP instructions=, CLAUDE.md, AI_INSTRUCTIONS.md) now mirror each other. No drift between what each client tells the AI.
• Setup → Maintenance state machine — AI detects placeholder vs real content via concrete signals, populates memory through MCP write tools (not direct file edits).

100% local. No exceptions.

• Memory lives at <your-project>/.projectmem/ (local to each project) and ~/.projectmem/global/ (machine-wide).
• Nothing uploads anywhere. No accounts. No telemetry. No "anonymous usage data."
• Team sharing is explicit: pjm global export > team-gotchas.json → commit → teammate runs pjm global import. You choose what to share and when.

Install

pip install projectmem

Requires Python 3.10+. MIT licensed.

Quick start

cd your-project
pjm init                           # creates .projectmem/, installs git hooks, starts watcher
# Open your AI client (Claude Desktop / Cursor / Antigravity / Codex)
# Ask anything about the project — memory is now wired

For the multi-project / monorepo setup, AI client configuration per tool, and the full feature tour, see the User Guide.

Upgrade notes

This is your first chance to install projectmem from real PyPI — there's nothing to upgrade from. If you previously installed from TestPyPI or from source, pip install --upgrade projectmem replaces it cleanly. Memory files (.projectmem/events.jsonl and global store) load without migration.

Acknowledgements

Built in the open, on a single MacBook, over six months of dogfooding. Every bug in the launch matrix was a real bug found by real use — not synthetic test data. Special thanks to the four MCP client teams whose work made this possible: Anthropic (Claude Desktop), Anysphere (Cursor), Google (Antigravity), and OpenAI (Codex).

The 46 lessons logged during this release are publicly tracked in report/LESSONS_AND_FIXES.md. Read them if you want to see how the sausage actually gets made.

---

Star the repo · Try it now · Read the guide · Report a bug