Persistent memory for AI coding agents

Your AI agent has amnesia

Every new session starts from scratch. You re-debug the same bugs, re-explain decisions you already made. ContextPool gives your agent persistent memory across sessions.

Commands

$ curl -fsSL https://raw.githubusercontent.com/syv-labs/cxp/main/install.sh | sh

$ cxp init claude-code --local

Found 14 Claude Code session(s) for this project.

Summarized 14 session(s) → 47 insight(s) extracted.

Top insights:

  • bug: ESM import fails silently in test runner
  • decision: chose SQLite over Postgres for local-only storage
  • gotcha: reqwest needs rustls-tls feature, not default openssl

✓ Your agent will now recall these automatically via MCP.

How it works

Step 1: Install

One curl command. Single static binary. No runtime dependencies. Runs on macOS, Linux, and Windows.

Step 2: Initialize

cxp scans your past Cursor and Claude Code sessions and extracts engineering insights using an LLM.

Step 3: Recall

Your agent automatically loads relevant past context via MCP at session start. No prompting needed.

What your agent remembers

Not conversation summaries. Actionable engineering knowledge distilled from your sessions.

Bugs & root causes

  • “Hydration error on Next.js when using Date.now() in SSR”

Fixes & solutions

  • “Use dynamic import() with { ssr: false } to skip hydration”

Design decisions

  • “API keys hashed with SHA-2 — never stored in plain text”

Gotchas & patterns

  • “reqwest needs rustls-tls feature, not default openssl”

Works where you work

Zero config in Claude Code. One JSON entry everywhere else.

Zero config

No API key needed. CXP uses your existing Claude Code authentication automatically. Just install and init.

$ cxp init claude-code

Built for reliable team memory

ContextPool keeps memory durable, portable, and safe by design.

Stable project IDs

Project IDs are derived from your git remote URL, so teammates resolve to consistent IDs automatically.

Multi-backend LLM routing

Provider fallback chain: Claude CLI → Anthropic API → OpenAI → NVIDIA for resilient extraction.

MCP protocol

Agents query memory via standard MCP tools. No custom integration glue is required.

Keychain storage

API keys are stored in the system keychain, with a safe file fallback where keychain is unavailable.

Team memory

Your teammate debugged this last week. Your agent should know that. Push insights to a shared pool. Everyone on the team pulls the collective knowledge.

$ cxp auth cxp_team_abc123

Authenticated as team "acme-eng" (free plan).

$ cxp pu

Pricing

Local mode is free forever. Cloud sync for teams.

Local

  • Free forever
  • Unlimited local insights
  • All IDE integrations
  • 4 LLM backends
  • Secret redaction
  • MCP server

Pro

Team access + cloud sync with a 7-day free trial.

  • $7.99/month
  • Everything in Local
  • Team access
  • Unlimited insights
  • Unlimited projects
  • 7-day free trial

Privacy first

Local-first by default: raw transcripts stay on your machine, while only extracted insights are synced when you opt in.

  • Stays local: Raw transcripts never leave your machine.
  • Redaction-first: Secrets are stripped before LLM processing and again before cloud sync.
  • To cloud (opt-in): Only extracted insights are synced, never full transcripts.

Stop re-discovering what you already know

Install in 30 seconds. Your agent starts remembering immediately.

Commands

$ curl -fsSL https://raw.githubusercontent.com/syv-labs/cxp/main/install.sh | sh

Or create a team account to sync with your teammates.