Stop your AI from forgetting your codebase. And from suggesting rm -rf.
DevMemory is a VS Code extension that gives Claude Code, Copilot, Cursor, Cline, Continue, and on-prem LLMs a persistent, auditable memory of your repo — and blocks destructive commands before you paste them in.
No telemetry by default. No network calls. Everything is plain Markdown under
.ai-memory/. Open the file,git diffit, delete it. Your machine, your files.
Works with: Claude Code · GitHub Copilot · Cursor · Cline · Continue · Codeium / Windsurf · on-prem LLMs (Llama, Bedrock, Azure OpenAI, …) — anything you can paste a prompt into.
What you actually get today
- Auto-generated
CLAUDE.md/AGENTS.md/.cursorrules/copilot-instructions.mdthat coexist with your hand-written rules. DevMemory only writes inside its markers; everything you write outside them is preserved on every export. - Inline paste guard. When you paste anything an AI gave you with
rm -rf,drop database,git push --force, fork bombs,format C:,dd if=/dev/zero of=/dev/sda,curl … | sh, or any of 20+ other destructive shapes, DevMemory blocks it modally before it lands in your editor and offers an Undo. Tests, docs, and.ai-memory/are exempt. - One-click resume prompt. Click Start AI Session in the sidebar; the extension reads your local memory, builds a fresh resume prompt, and copies it to your clipboard. Paste into any chat. No re-explaining the project.
- Session capture, validated. End AI Session copies a structured wrap-up prompt; you paste the AI's reply back, and DevMemory validates the response (
rm -rf, simulated content, missing files, empty summaries) before appending the session log and refreshing yourcurrent-state.mdandtasks/next-actions.md. - Local memory health check. Audits the
.ai-memory/store for placeholders, missing files, and sessions saved with warnings. Writes a plain-Markdown report. - Optional local-only telemetry. Off by default. When you flip
devmemory.telemetry.enabledon, DevMemory appends events (durations, command names, paste-guard hits) to.ai-memory/telemetry/events-YYYYMMDD.jsonl. Inspect, export, or wipe it with one click. Never sent over the network.
The sidebar adapts to where you are
The hero CTA at the top of the sidebar changes automatically as your memory progresses — Scan this project → Teach the AI → Save what the AI told me → Resume an AI session → Save what the AI did. No more digging through 17 menu items to find what's next.
Three pains it solves
| Pain | Without DevMemory | With DevMemory |
|---|---|---|
| The AI forgets your architecture every session | Re-explain the codebase. Hope it doesn't hallucinate. | Click Resume AI Session → prompt with your real architecture lands on clipboard, or use @devmemory /resume in chat. |
| The AI invents files / functions / decisions | Catch it in code review (or not). | Saving the session triggers a validator that flags simulated content and references to missing files. |
The AI proposes rm -rf or DROP DATABASE |
Read carefully before pasting. Hope. | DevMemory blocks the paste modally. One click to undo. |
How DevMemory compares
DevMemory does not replace the native chat in Claude Code, Cursor, or Copilot. It produces and audits the context you feed into them — and adds the safety net those tools don't.
| Copilot | Cursor | Cline / Continue | Cody | DevMemory | |
|---|---|---|---|---|---|
| Local-only memory | – | – | – | – | ✓ |
| Inline destructive-paste guard | – | – | – | – | ✓ |
| Multi-AI portable memory (CLAUDE.md / AGENTS.md / cursorrules / copilot) | – | – | – | – | ✓ |
| Audit trail of AI replies | – | – | – | – | ✓ (Markdown + signed log) |
| Telemetry off by default | – | – | – | – | ✓ |
| Open core (FSL → Apache 2.0) | – | – | partial | – | ✓ |
Already maintaining a hand-written CLAUDE.md and you're happy? Keep it. Run DevMemory: Export AI Context Files and DevMemory will only manage the block between its markers — your handwritten rules stay untouched.
Install
code --install-extension devmemory-ai.devmemory-ai-vscode
A DevMemory AI icon appears in the Activity Bar. Open it and the sidebar always shows your next recommended action.
For local development / building from source:
npm install
npm run package:vscode
code --install-extension apps/vscode-extension/devmemory-ai-vscode-0.3.0.vsix --force
Five-step flow
A native walkthrough opens on first install (or anytime via DevMemory AI: Open Welcome Walkthrough) so you can do all five steps without leaving the editor:
- Scan This Project — scans the workspace (with strict, security-aware exclusions) and creates
.ai-memory/. - Teach The AI About This Project → Save What The AI Told Me — bootstraps
project-summary,architecture,current-state, andnext-actionsusing your AI of choice (clipboard or@devmemorychat). - Resume AI Session — copies a fresh resume prompt that primes the AI with current memory.
- Wrap Up Session → Save What The AI Did — appends a session log (validated). Updates
current-state.mdandtasks/next-actions.md. - Check Memory Health — audits the local store and writes
.ai-memory/health-check.md.
For security-conscious teams
DevMemory was originally built for fintech, healthtech, govtech, defense, and consultancies under NDA. The core posture stays the same:
- No network calls. No telemetry by default. Verifiable in source — there is no HTTP client and no analytics dependency in the extension.
- Workspace-only writes. All output is inside
.ai-memory/. Commit it, ignore it, archive it — your call. - ~140 default exclude patterns for secrets, credentials, build artifacts, virtual envs, signing keys, Terraform state, mobile config, IDE caches.
- Workspace trust required before any scan. Confirmation modal before initial setup.
- Audit trail in plain Markdown. Diff it. Review it. No opaque indices.
- Optional signed audit pack (coming in v0.3): one-click export of every session log and every paste-guard event, signed with a project-local ed25519 key.
See PRIVACY.md for the full statement.
Talk to DevMemory from inside chat — @devmemory
Type @devmemory /resume inside Copilot Chat or any VS Code Chat host and DevMemory injects a fresh resume prompt directly into the chat. No clipboard, no context switch.
| Subcommand | What it does |
|---|---|
/resume |
Streams the current resume prompt into chat. |
/decisions |
Shows your decisions log. |
/architecture |
Shows the architecture summary. |
/save-session <pasted-AI-response> |
Validates the response (destructive commands, simulation, missing files) and saves it. |
DevMemory does not call a model itself — the participant only feeds local context. Your AI host stays in charge of which model answers.
Connect any MCP-aware client (Claude Code, Cursor, Cline, Continue)
Run DevMemory AI: Enable MCP Server and DevMemory writes .vscode/mcp.json (managed block — your other entries are preserved). Claude Code / Cursor / Cline / Continue auto-discover four tools:
memory.read— read a section (project-summary,architecture,current-state,next-actions,decisions,bugs).memory.search— search decisions and bugs logs.memory.audit— run the destructive-command validator on any text.memory.appendSession— save a session draft from inside the agent.
The server is pure stdio JSON-RPC. Never opens a network port. Opt-in via devmemory.exposeMcpServer.
What's not here yet
- E2E-encrypted memory sync for teams (server-side; opt-in) — on the v0.5 roadmap.
- JetBrains and Neovim ports — planned, not started.
- Cursor as a first-class chat host —
@devmemoryworks in Copilot Chat today; Cursor support depends on Cursor's chat-participant support.
If any of these are blockers, file an issue. We prioritize what real users need.
License
Functional Source License 1.1, with an Apache 2.0 future grant (FSL-1.1-Apache-2.0). Free for any internal, educational, and non-competing commercial use. Auto-converts to Apache 2.0 on 2028-04-28.
Reporting issues
https://github.com/jukafarneiy/devmemory-ai/issues — please include .ai-memory/health-check.md for any "memory looks weird" reports.