Open Source · MIT Licensed
Your AI has
amnesia.
Persistent memory and cognitive context for AI. brain-mcp turns your conversation history into a queryable brain that any MCP client can use.
25 MCP tools · millisecond recall · 100% local · works with Claude, Cursor, Windsurf
The only command you need to remember
“use brain”
Type it in any conversation with Claude, Cursor, or Windsurf. Your AI instantly gains access to your full thinking history — every decision, every open question, every domain you've worked in.
Before & After
❌ Without brain-mcp
“I was working on something with embeddings last month... let me check my chat history... which conversation was it... okay I found three but I can't tell which one had the decision about chunking strategy...”
~30 minutes of context fumbling. Maybe 60% recovered.
✅ With brain-mcp
> tunnel_state("my-project")
🧠 Tunnel State: my-project
Stage: executing | Tone: focused
Open questions: 3 | Decisions: 5
❓ Top open:
- Which chunking strategy to use?
- A/B test results pending
✅ Recent decisions:
- Use semantic chunking over fixed
- Target 768-dim embeddings
⏱️ Instant recall. Full context.What It Does
State Recovery
Pick up exactly where you left off in any project or domain. Like loading a saved game — open questions, decisions, current stage, all recovered instantly.
Thought Synthesis
Your AI synthesizes what you think about a topic from your real conversations. Not what you bookmarked — what you actually said and decided.
Attention Protection
Know the cost of switching contexts before you switch. Catch dropped threads before they disappear. Surface dormant work you forgot about.
Semantic Search
Find conversations by concept, not keyword. Local embeddings via nomic-embed-text — no API key needed, no data leaving your machine.
25 MCP Tools
8 cognitive prosthetic tools + 17 generic tools. All exposed via Model Context Protocol.
🧠 Cognitive Prosthetic (8)
tunnel_stateReconstruct your save-state for any domain. Like loading a saved game.
context_recoveryFull re-entry brief when returning to a dormant topic.
switching_costQuantified cost of switching between domains.
open_threadsGlobal unfinished business across all your domains.
dormant_contextsAlarm for abandoned topics with unresolved questions.
trust_dashboardSystem-wide proof the safety net works.
cognitive_patternsWhen and how you think best, backed by data.
tunnel_historyEngagement meta-view for any domain over time.
🔧 Generic (17)
semantic_searchsearch_conversationsunified_searchsearch_summariessearch_docsunfinished_threadswhat_do_i_thinkalignment_checkthinking_trajectorywhat_was_i_thinkingget_conversationconversations_by_datebrain_statsgithub_searchquery_analyticslist_principlesget_principle6 search · 4 synthesis · 3 conversation · 1 GitHub · 1 analytics · 2 meta. Full reference →
How It Compares
| brain-mcp | Mem0 | Letta | |
|---|---|---|---|
| Memory model | Conversation archaeology | Key-value facts | Tiered memory |
| State recovery | ✅ 8 prosthetic tools | ❌ | ❌ |
| Data source | Your existing conversations | Runtime extractions | Agent conversation |
| Runs where | 100% local | Cloud or self-host | Cloud or self-host |
| Domain modeling | ✅ Multi-domain tracking | ❌ | ❌ |
| Context switching | ✅ Switching cost analysis | ❌ | ❌ |
| Protocol | MCP (any client) | REST API | REST API |
| Price | Free (MIT) | Freemium | Freemium |
Most memory systems store extracted facts. brain-mcp reconstructs cognitive state — where you were in a problem, what you'd decided, and what questions are still open.
Why This Exists
Built by someone with ADHD who kept losing their train of thought. Not files — the state of what they were reasoning about. The decisions already made. The questions still open. Gone on every context switch.
Open source because this problem isn't unique. Whether you have ADHD or you just come back to a project after a week and spend the first hour re-figuring out what you already knew — brain-mcp is for you.
Context that survives is context that compounds.
Your AI conversations already contain everything you need. brain-mcp makes it queryable.