cortexmem
Persistent memory for AI coding agents. Builds semantic memory from git history and codebase, searchable via MCP tools.
README
cortexmem
Persistent memory for AI coding agents. Zero config, works with Cursor, Claude Code, Codex, and any MCP-compatible editor.
AI coding agents lose all context when a session ends. CortexMem fixes this by building a semantic memory store from your git history, codebase, and session context, then making it searchable via MCP tools.
Setup
Step 1: Initialize your project
cd your-project
npx cortexmem init
This scans your git history and codebase, embeds everything locally, and stores it in .cortexmem/store.db. It also generates editor config files (CLAUDE.md, .cursorrules, codex.md) that instruct AI agents to use cortexmem automatically.
First run downloads the embedding model (~30MB, one-time). Subsequent runs are incremental and only re-index new commits and changed files.
$ npx cortexmem init
CortexMem — initializing context for /Users/you/my-project
Full scan — first-time initialization...
Found 142 commits → 87 chunks
Found 38 files → 52 chunks
Embedding 139 chunks...
Storing in database...
Building project summary...
Generating editor configs...
Created: CLAUDE.md, .cursorrules, codex.md
Done!
Summary:
Git commits indexed: 142
Source files scanned: 38
Total chunks stored: 139
Storage: /Users/you/my-project/.cortexmem/store.db
Add to your MCP config to start using cortexmem with your AI agent.
You can optionally include a project spec or requirements doc:
npx cortexmem init ./PROJECT.md
Step 2: Add to your editor's MCP config
Cursor (add to ~/.cursor/mcp.json):
{
"mcpServers": {
"cortexmem": {
"command": "npx",
"args": ["-y", "cortexmem"]
}
}
}
Claude Code (add to ~/.claude.json or project settings):
{
"mcpServers": {
"cortexmem": {
"command": "npx",
"args": ["-y", "cortexmem"]
}
}
}
With LLM-powered compaction (optional, add your Anthropic API key):
{
"mcpServers": {
"cortexmem": {
"command": "npx",
"args": ["-y", "cortexmem"],
"env": {
"ANTHROPIC_API_KEY": "sk-ant-..."
}
}
}
}
Restart your editor. CortexMem is running.
ANTHROPIC_API_KEYis optional. It enables LLM-based session compaction viasummarize_session. Without it, everything else works and compaction uses a deterministic fallback.
Step 3: There is no step 3
The generated editor config files (CLAUDE.md, .cursorrules, codex.md) instruct your AI agent to use cortexmem automatically. It will:
- Load context from previous sessions on startup
- Save decisions, discoveries, and constraints as you work
- Compact memory at session end
No manual tool calls needed.
Example: What a session looks like
Session 1: You start working on auth
Your AI agent automatically calls get_context at session start:
## CortexMem Context — my-project
Initialized: 2026-03-08T10:30:00Z
### Project Overview
my-project: Node.js/TypeScript API server. 142 commits, 38 files.
Stack: Express, PostgreSQL, Jest. Main modules: auth, payments, users.
### Index Stats
- Commit Summaries: 87 chunks
- Code Summaries: 52 chunks
During work, the agent saves context automatically:
save_context({
context_type: "decision",
content: "Using JWT with refresh tokens for auth. Access tokens expire in 15min, refresh tokens in 7 days. Stored in httpOnly cookies, not localStorage.",
related_files: ["src/auth/jwt.ts", "src/middleware/auth.ts"]
})
→ Saved decision context (id: 12, session: a1b2c3, branch: main)
save_context({
context_type: "constraint",
content: "Auth middleware must never be modified directly. Extend via plugins in src/auth/plugins/",
related_files: ["src/middleware/auth.ts"]
})
→ Saved constraint context (id: 13, session: a1b2c3, branch: main)
save_context({
context_type: "state",
content: "Auth implementation: JWT service done, middleware done, refresh token rotation TODO",
related_files: ["src/auth/jwt.ts"]
})
→ Saved state context (id: 14, session: a1b2c3, branch: main)
At session end, the agent calls summarize_session:
summarize_session({ session_summary: "Implemented JWT auth with refresh tokens" })
→ Compaction complete:
Session: Compacted 3 entries into session summary
Branch (main): Updated branch summary
Project: Updated project overview
Session 2: Different day, context is preserved
The agent calls get_context and immediately has full context:
## CortexMem Context — my-project
### Project Overview
my-project: Node.js/TypeScript API with JWT auth (access + refresh tokens),
PostgreSQL, Express. Auth module complete, payment refactor in progress.
### Branch: main
JWT auth implemented with httpOnly cookies. Auth middleware uses plugin
architecture (never modify directly). Refresh token rotation still TODO.
### Recent Sessions (main)
#### Session a1b2c3 (2026-03-08)
Implemented JWT authentication with refresh tokens. Access tokens expire
in 15min, refresh in 7 days. Created plugin-based auth middleware.
Refresh token rotation is the next task.
### Index Stats
- Decisions: 1 chunks
- Constraints: 1 chunks
- State: 1 chunks
- Commit Summaries: 87 chunks
- Code Summaries: 52 chunks
The agent can also search for specific context:
get_context({ query: "auth middleware", depth: 3 })
→ ## CortexMem Context — my-project
Query: "auth middleware" | depth: 3
### [project > branch:main > session:a1b2c3] (87% match)
JWT auth with refresh tokens. Plugin-based middleware architecture.
**Details:**
- [Constraint] Auth middleware must never be modified directly. Extend via plugins
- [Decision] Using JWT with refresh tokens for auth. Access tokens expire in 15min...
Re-running init (incremental)
When you come back after more commits:
$ npx cortexmem init
CortexMem — initializing context for /Users/you/my-project
Incremental update — scanning changes since last init...
8 new commits → 6 chunks
3 files changed
3 changed files → 4 chunks
Embedding 10 chunks...
Storing in database...
Building project summary...
Done!
Summary (incremental):
Git commits indexed: 8 (new)
Source files scanned: 38
Total chunks stored: 10 (new)
How It Works
cortexmem initscans your git history and codebase, chunks and embeds everything locally- Everything is stored in
.cortexmem/store.db, a single SQLite file portable across editors and machines - Your AI agent uses 4 MCP tools to search, save, and compact context
- Context is organized in a pyramid: project, branch, and session summaries with raw chunks underneath
The Context Pyramid
Project Summary ← "What is this project about?"
├── Branch: main ← "What's happening on main?"
│ ├── Session a1b2c3 ← "What did we do 2 days ago?"
│ └── Session d4e5f6 ← "What did we do yesterday?"
└── Branch: feature/payments ← "What's the payments work?"
└── Session g7h8i9
get_context()returns the pyramid overview (~500-800 tokens)get_context({ query: "..." })searches hierarchically, matching summaries first and drilling into raw chunks only when neededsummarize_session()rolls up: session chunks → session summary → branch summary → project summary
What Gets Indexed
| Source | What's Extracted |
|---|---|
| Git log | Commit messages, descriptions, file change patterns |
| Source files | Code structure, functions, classes, patterns |
| Config files | Stack, tooling, dependencies |
| Docs (.md) | Documentation content |
| Project file | Specs, requirements (via cortexmem init <file>) |
| Session context | Decisions, constraints, discoveries saved by the agent |
MCP Tools
| Tool | When to use | What it does |
|---|---|---|
get_context |
Session start, or when you need specific context | Returns pyramid overview (no args) or hierarchical search (with query). Depth 0-3 controls granularity. |
save_context |
When the agent makes a decision, discovers something, notes a constraint | Embeds and stores instantly. Types: decision, constraint, state, discovery, preference. |
summarize_session |
End of session | Compacts saved context into the pyramid. Uses Claude Haiku if ANTHROPIC_API_KEY is set, deterministic fallback otherwise. |
get_status |
Anytime | Quick stats: chunk counts by type, storage location, last init time. |
Context Types
| Type | Purpose | Example |
|---|---|---|
| decision | Architectural/technical choices | "Chose PostgreSQL over MongoDB for ACID transactions" |
| constraint | Hard rules to never violate | "Never modify auth middleware directly" |
| state | Current WIP status | "Payment refactor: 2/4 services done" |
| discovery | Non-obvious codebase facts | "UserService is called from 6 places, not 3" |
| preference | Code style conventions | "Snake_case for variables, PascalCase for classes" |
CLI Commands
cortexmem init [project-file] Scan git history + codebase, build context store
Incremental on re-run, only indexes new changes
cortexmem inject <file> Inject/update a project file (spec, requirements)
cortexmem status Show what's stored
cortexmem Start MCP server (used by AI editors)
Portability
CortexMem stores everything in a single file: .cortexmem/store.db
# Move to a new machine
scp .cortexmem/store.db user@newmachine:~/project/.cortexmem/
# Share with teammates (commit it)
git add .cortexmem/store.db
# Switch editors, same file works everywhere
# Claude Code -> Cursor -> Codex, no migration needed
Environment Variables
| Variable | Purpose | Default |
|---|---|---|
ANTHROPIC_API_KEY |
Enables LLM compaction in summarize_session |
none (deterministic fallback) |
CORTEXMEM_MAX_TOKENS |
Default max tokens for get_context |
3000 |
CORTEXMEM_MODEL |
Model for compaction | claude-haiku-4-5-20251001 |
Architecture
- Embeddings:
all-MiniLM-L6-v2via@xenova/transformers. Runs locally, no API key needed, ~30MB model - Storage: SQLite via
sql.js(WASM). Zero native dependencies, works on any OS - Search: Hybrid keyword + vector search. Keywords by default, vector when model is warm. Both work offline.
- Transport: MCP stdio. Works with any MCP-compatible editor
Development
git clone https://github.com/Ashprakash/cortexmem.git
cd cortexmem
npm install
npm test # run 106 tests
npm run dev # run with tsx
npm run build # compile TypeScript
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.