coretx 0.9.5

Structured memory for AI that persists, attributes, and connects

CoreTx Connect

Give your AI assistant persistent memory that survives across conversations, sessions, and devices.

CoreTx Connect is a lightweight MCP server (~450 lines, zero dependencies) that stores knowledge objects (KOs) via the CoreTx API. Your assistant captures decisions, facts, preferences, and insights as you work, and retrieves them automatically in future sessions.

What changes: Your AI remembers what you discussed last week. It recalls architectural decisions from three months ago. It knows your preferences without being told again. When context compaction would normally erase critical details, KOs preserve them.

Quick Start

Claude Desktop (automatic setup)

# 1. Install
pip install coretx

# 2. Run setup (writes Claude Desktop config automatically)
coretx setup <your-api-key>

# 3. Restart Claude Desktop

Claude Code (CLI)

# 1. Install
pip install coretx

# 2. Add to Claude Code
claude mcp add --transport stdio ko-memory -- coretx serve

# 3. Set your key (add to your shell profile for persistence)
export CORETX_API_KEY=npub_sk_xxx

# 4. Verify everything works
coretx --test

The --test command checks API connectivity, writes and reads a test KO, and confirms your setup is working.

Getting an API key

You can create your own API key at sapience.coretx.ai under API Keys, or use coretx init to create an account and key from the terminal.

How It Works

When you start a conversation, your assistant calls ko_context to load relevant knowledge from previous sessions. As you work, it proactively captures decisions, facts, and insights via ko_write. These KOs persist server-side and are available in every future session, on any device, with any AI provider.

You: "Let's use PostgreSQL on Supabase for the database"
  -> ko_write(subject='database', predicate='decided as', object_value='PostgreSQL on Supabase')

[Three weeks later, new session]

Assistant calls ko_context -> loads the database decision automatically
You: "What database are we using?"
  -> Already knows: PostgreSQL on Supabase

Tools

Cloud tools (persistent, require API key)

Tool	What it does
ko_write	Store a knowledge object. Called proactively by the assistant.
ko_read	Retrieve a specific KO by subject + predicate.
ko_search	Search KOs by keyword or natural language.
ko_list	List recent KOs, optionally filtered by project.
ko_delete	Delete a KO by subject + predicate.
ko_context	Load session-start context (pinned + recent + relevant KOs).
ko_stats	Get store statistics (total KOs, projects, pinned count).
ko_share	Share KOs with another CoreTx user by email or ORCID.

Local tools (session-scoped, no API key needed)

Tool	What it does
ko_set_context	Set project/goal for smarter retrieval.
ko_working_memory	Ephemeral scratch memory (auto-expires after 1 hour).
ko_todos	Task tracking within the session.

CLI Commands

Bootstrap (cold start from Claude Code history)

If you have existing Claude Code sessions, bootstrap your knowledge store from them:

coretx bootstrap --dry-run    # preview what would be extracted
coretx bootstrap              # extract and write KOs
coretx bootstrap --path /custom/logs/path

This scans your Claude Code JSONL logs (~/.claude/projects/) and extracts decisions, facts, problems, and preferences as KOs. Projects are inferred from directory names.

Reclassify (organize unassigned KOs)

If you have KOs without a project, classify them based on your existing project keywords:

coretx reclassify --dry-run   # preview classifications
coretx reclassify             # apply classifications

Other commands

coretx --test      # verify API connectivity
coretx --version   # show version
coretx setup <key> # configure Claude Desktop

Environment Variables

Variable	Required	Default
CORETX_API_KEY	Yes (for cloud tools)	--
CORETX_API_URL	No	https://api.coretx.ai
CORETX_AUTOSEED	No	1 — set to 0 to disable the session-start project snapshot
CORETX_INIT_CONTEXT	No	1 — set to 0 to disable the MCP-initialize memory preamble

Warm session startup

Two layers work together so sessions open with context automatically, without the assistant needing to call any ko-memory tool:

1. MCP initialize preamble — when your MCP client starts the server, the initialize response carries your pinned + recent semantic KOs across all projects, embedded directly in the protocol's instructions field. The assistant reads this as part of session setup. Controlled by CORETX_INIT_CONTEXT.

2. First-tool-call project snapshot — once a project is resolvable (from ko_set_context or the basename of cwd), the server prepends a project-scoped block of pinned KOs, open problems, and recent decisions to that tool's result. Fires at most once per session. Controlled by CORETX_AUTOSEED.

Both are best-effort: on timeout, API error, or when CORETX_API_KEY is unset, the server falls back to static instructions and the session still starts cleanly.

Pinned is protected from truncation. Each preamble is capped at ~8KB, but the Pinned section is always included in full — even if it alone overruns the budget. Only non-pinned sections (semantic / recent) get trimmed when context is large. This matches the user contract: pinning a KO means "always surface this."

Capture control (privacy)

Knowledge capture is paused by default at session start and hard-enforced at the server layer: any ko_write attempt while paused is rejected without touching the network, regardless of what the model tries to do.

• Code mode — as soon as ko_set_context yields a resolvable project (from explicit project or cwd basename), capture auto-enables. Users running Claude Code in a project directory notice no change in behavior.
• Chat mode — without a project, capture stays paused. The model is instructed to tell users "memory capture is off for this session" and call ko_resume only when the user explicitly asks to enable it.
• Explicit pause/resume locks the state. Once a user calls ko_pause or ko_resume, later ko_set_context calls don't silently flip it. This is the privacy-critical invariant: the user's choice wins.

Three tools manage this:

Tool	Use
ko_pause	Block all writes for this session. Optional reason surfaced on status checks. Use when the user describes a sensitive topic (medical, legal, financial, family, therapy) or explicitly asks not to be remembered.
ko_resume	Re-enable writes. Use when the user confirms they want persistent memory in a Chat session, or after a pause.
ko_status	Report the current state (paused/active, default/explicit, project context).

Troubleshooting

"CORETX_API_KEY not set" — Make sure your API key is set via the setup command or in your environment.

Server not appearing in Claude Desktop — Restart Claude Desktop after running setup.

"HTTP 401" errors — Your API key is invalid or expired. Regenerate one at sapience.coretx.ai under API Keys.

Wrong Python version — Re-run coretx setup <key> — it uses the correct Python automatically.

Team Setup

Each team member needs their own API key. KOs are per-user and private by default. To share knowledge across the team, use ko_share with a colleague's email or ORCID.

# Each team member runs:
pip install coretx
claude mcp add --transport stdio ko-memory -- coretx serve
export CORETX_API_KEY=npub_sk_xxx  # their own key
coretx --test

Architecture

• Local server: MCP over stdio for Claude Code / Desktop
• Remote server: Streamable HTTP for claude.ai browser (coretx-remote)
• Storage: All persistent state at api.coretx.ai (no local database)
• Auth: Bearer token (npub_sk_* API keys), OAuth 2.1 for browser
• Signing: Ed25519 client-side signing (local server only)
• Server name: ko-memory