memex-chats 0.2.3

Local MCP server that indexes your Claude.ai chats and your Claude Code sessions, and exposes them to Claude Code (stdio) and to Claude.ai (remote MCP connector).

Memex

Your Claude.ai chats and your Claude Code sessions live in two worlds that never talk. Memex makes them one memory you can search from both.

Local-first MCP server that indexes your Claude.ai chat history and your local Claude Code / terminal sessions, and exposes them to Claude Code (stdio) and Claude.ai (remote MCP connector).

Status: alpha, 0.2.3 on PyPI. Phases 0 to 4 and 6 closed, security-audited (plus four adversarial red-team rounds). Shipped: hybrid search, live capture, auto-summaries, chat ↔ repo association, the Claude.ai remote connector (GitHub OAuth), and Claude Code / terminal ingestion with secret redaction and automatic sync.

End-to-end demo: a chat on claude.ai captured seconds earlier by the Chrome extension, recalled from Claude Code via list_recent_chats and get_chat. No extra prompting, no manual context handoff.

The problem

Brainstorming and planning happen in Claude.ai. Execution happens in Claude Code. The two worlds do not talk to each other: Claude Code cannot read a chat of yours from Claude.ai, not even the one that originated the task it is currently working on. The memory Anthropic shipped on Claude.ai (March 2026) is curated, not full history, and lives isolated inside Claude.ai.

Memex fills that gap: runs locally, indexes the entire corpus of your chats, and exposes them as MCP tools so Claude can search and pull past context whenever it needs to.

How it works

[Claude.ai]
    ↓  (official JSON export / Chrome ext)
[Ingestor]  →  [SQLite + sqlite-vec]  →  [local embeddings (fastembed / Ollama)]
                                    ↓
                          [core: storage + retrieval]
                                    ↓
                  [MCP stdio]  ───→  Claude Code, Claude Desktop
                  [MCP SSE/HTTP] ──→ Claude.ai (coming soon)

Design: pure core (storage, ingest, embeddings, retrieval) decoupled from transport. The same engine serves both stdio and remote MCP without a rewrite.

Installation

Memex is a Python package. You do not need to install Python yourself: the installer below uses uv, which downloads the right Python (3.13) and all dependencies into a local environment. The only thing you need beforehand is git.

Embeddings work out of the box: by default Memex uses fastembed, which downloads a quantized 130 MB model the first time you ingest. No Ollama, no API key, no extra setup. (Ollama is optional, see Embeddings backend below.)

Note: a PyPI install gives you the CLI and the MCP servers, but the live-capture Chrome extension, the autostart scripts, and the launchd/systemd templates only come with the repo, so installing from source (below) is still the recommended path.

macOS / Linux

git clone https://github.com/dioniipereyraa/memex
cd memex
./scripts/install.sh

install.sh installs uv if you do not have it, runs uv sync, and verifies the install with memex doctor. That is the whole setup.

Windows (PowerShell)

git clone https://github.com/dioniipereyraa/memex
cd memex
powershell -ExecutionPolicy Bypass -File .\scripts\install.ps1

Same three steps as the macOS script (install uv, sync, verify). If git is not installed, get it from git-scm.com first.

Manual (any OS, if you prefer explicit steps)

# 1. Install uv (skip if you already have it):
#    macOS/Linux:  curl -LsSf https://astral.sh/uv/install.sh | sh
#    Windows:      powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
git clone https://github.com/dioniipereyraa/memex
cd memex
uv sync               # installs Python 3.13 + dependencies into .venv
uv run memex doctor   # verify

First run

# 1. Request your official Claude.ai export (Settings → Privacy → Export data),
#    then ingest it (first run downloads the embedding model, ~30s):
uv run memex ingest /path/to/your-export.zip

# 2. Optionally index your local Claude Code / terminal sessions:
uv run memex ingest-claude-code

# 3. Search, or wire it into Claude Code (see "Wiring it into Claude Code"):
uv run memex search "your query" -n 5
uv run memex stats

(If you installed with a script, the commands above work as written. If you installed the published PyPI package instead, drop the uv run prefix.)

Embeddings backend

Default is fastembed (zero-config). To route through a local Ollama instead (e.g. you already run it for other models):

export MEMEX_EMBED_BACKEND=ollama
ollama pull nomic-embed-text          # install Ollama first: https://ollama.com/download

On Windows, install Ollama from ollama.com/download before setting MEMEX_EMBED_BACKEND=ollama. This is optional; fastembed needs none of it.

Diagnostics

Run memex doctor any time something is not working. It checks Python version, database, embedder, live-capture server, summarizer config, registered repos, and indexed corpus. Reports OK / WARN / FAIL per check.

MCP server tools (v1)

• search_chats(query, limit=5, source?, mode="hybrid", repo?) searches the corpus. Modes: hybrid (default, combines vector search + FTS5 BM25 via Reciprocal Rank Fusion), semantic (vectors only), lexical (FTS5 only, ideal for proper nouns or exact terms). source filters by origin (conversations, design_chat, memory). repo boosts results associated to a registered repo (see "Repo associations" below). Deduplicated per conversation.
• get_chat(uuid, messages_limit=10, messages_offset=0) fetches a conversation with its messages, paginated. raw_content is omitted; each message is truncated to 1500 chars to stay inside the client's token budget (worst-case response ~17k chars). Long chats are paginated with messages_offset; max messages_limit is 100.
• list_recent_chats(limit=10, source?) lists the latest chats ordered by last update.
• find_related(context, limit=5, repo?) takes free-form text (a paragraph, a file contents, the current discussion) and returns chats that are semantically related. Pure vector search, no FTS, capped at 4000 input chars. Useful when you want "more like this" without typing a keyword query.

Search is also reachable from the CLI with memex search "query" --mode {hybrid|semantic|lexical}. For databases created before the hybrid FTS5 work, run memex reindex-fts once to populate the lexical index.

Wiring it into Claude Code

Once your local database is populated (memex ingest), start the MCP server with uv run memex-mcp. For Claude Code to discover it automatically, add a .mcp.json file at the root of your project (or a user-level server in ~/.claude.json):

{
  "mcpServers": {
    "memex": {
      "command": "uv",
      "args": ["run", "memex-mcp"],
      "cwd": "/absolute/path/to/the/memex/repo"
    }
  }
}

Set cwd to the absolute path where you cloned Memex (where pyproject.toml lives). Restart Claude Code and the tools search_chats, get_chat, list_recent_chats will show up in the session.

The same searches are also available from the CLI via uv run memex search "..." if you prefer them outside Claude Code.

Connecting from claude.ai (remote MCP, Phase 4)

memex serve-remote exposes the same 4 tools as a custom connector for claude.ai. One connector works across claude.ai web, Claude Desktop, and the mobile apps: the connection is brokered by your Claude account and always originates from Anthropic's cloud, never from your device. That has two consequences:

1. The server must be reachable on a public HTTPS URL (a hostname with a public IPv4 A record; localhost and private IPs are rejected by design). The intended setup is a tunnel that terminates TLS and proxies to Memex on loopback.
2. Auth is either none or full OAuth 2.0 (claude.ai has no field to paste a token). Memex uses OAuth backed by a GitHub OAuth App, restricted to an allow-list of GitHub usernames: anyone else who completes the OAuth dance still gets a 401 on every request. Your machine must be on (and the tunnel up) for the connector to respond.

1. Publish a URL with Tailscale Funnel

Install Tailscale, log in, then:

tailscale funnel --bg 8377

This prints your public URL, e.g. https://my-mac.my-tailnet.ts.net, TLS included. Any other tunnel works the same way (ngrok, Cloudflare Tunnel); Memex only needs the resulting https:// URL.

2. Create a GitHub OAuth App

At github.com/settings/developers > OAuth Apps > New OAuth App:

• Homepage URL: your funnel URL.
• Authorization callback URL: <funnel URL>/auth/callback.

Copy the Client ID and generate a Client Secret.

3. Configure and run

In .env (see .env.example):

MEMEX_REMOTE_BASE_URL=https://my-mac.my-tailnet.ts.net
MEMEX_GITHUB_CLIENT_ID=Iv1.xxxxxxxxxxxx
MEMEX_GITHUB_CLIENT_SECRET=xxxxxxxxxxxxxxxxxxxx
MEMEX_REMOTE_ALLOWED_GITHUB_LOGINS=your-github-username

uv run memex serve-remote   # listens on 127.0.0.1:8377, endpoint at /mcp

The server refuses to start with incomplete config (no insecure defaults: an empty allow-list would expose your whole chat history to any GitHub account).

4. Add the connector in claude.ai

Settings > Connectors > Add custom connector, with URL https://my-mac.my-tailnet.ts.net/mcp. claude.ai registers itself via dynamic client registration, sends you through the GitHub authorization, and the tools appear in your chats. OAuth state is persisted encrypted on disk, so restarting serve-remote does not break the connection.

Security model in short: loopback bind + tunnel, OAuth proxy over GitHub with a username allow-list enforced on every request (revoking the app on GitHub locks it out immediately), TrustedHostMiddleware pinned to the public hostname, and the same indirect-prompt-injection envelope on tool results as the stdio transport.

Indexing Claude Code / terminal sessions (Phase 6)

Memex also indexes your local Claude Code and terminal sessions, so a single search covers both halves of your history: what you discussed on claude.ai and what you built with Claude Code. Claude Code records each session as a JSONL file under ~/.claude/projects/; Memex reads them directly (no extension, no network).

uv run memex ingest-claude-code        # scans ~/.claude/projects/**/*.jsonl
uv run memex ingest-claude-code --path /custom/path   # alternate root

These sessions are stored under the claude_code source, searchable like any other chat (and filterable with source="claude_code"). Details:

• Incremental. Unchanged sessions are skipped (by content hash), so re-running after more work is cheap. Sessions grow append-only, so a re-scan picks up new turns and new sessions. Re-run it whenever you want to refresh, or wire it to a schedule.
• Repo-aware for free. Every session carries its working directory, so each is auto-associated with the registered repo of that cwd (run memex repos add <path> first). search_chats(repo=...) then boosts the sessions where you worked on that project.
• What is indexed: your prompts, the assistant replies, and tool calls as [tool_use: ...] / [result] markers. Excluded: assistant internal reasoning (thinking), parallel sub-agent side threads, and CLI plumbing (slash-command echoes, bash wrappers). Everything stays in the local SQLite DB.
• Secret redaction. Session logs capture real terminal output and file contents, which often contain credentials. Before storing, Memex masks common secret shapes (API keys, bearer/JWT tokens, PEM private keys, KEY=/SECRET= assignments, URLs with embedded passwords) as [REDACTED:...], and does not persist the raw block content for this source. Best-effort, not a guarantee: it catches well-known formats so third-party credentials that were never really part of a conversation stay out of the index (this matters because the remote connector can surface indexed text to claude.ai).

Automatic sync (keep Claude Code sessions fresh)

memex ingest-claude-code is manual. To index sessions automatically, use two complementary mechanisms (both run in the background at low priority, and the embedding model only loads when there is something new, so they barely cost anything):

1. SessionEnd hook — ingests a session the moment it closes (no delay, nothing lost). Add to ~/.claude/settings.json:

   {
     "hooks": {
       "SessionEnd": [
         { "matcher": "*", "hooks": [
           { "type": "command", "command": "/absolute/path/to/memex/scripts/session-end-hook.sh" }
         ] }
       ]
     }
   }
   

   The hook reads the closed session's transcript_path and ingests just that one session, detached, returning immediately (SessionEnd hooks cannot block Claude Code).

2. Periodic backstop — a scheduled scan that catches anything the hook missed (e.g. a crash). On macOS, with launchd:

   sed "s|__REPO__|$(pwd)|g" scripts/com.memex.ingest-claude-code.plist.template \
     > ~/Library/LaunchAgents/com.memex.ingest-claude-code.plist
   launchctl load ~/Library/LaunchAgents/com.memex.ingest-claude-code.plist
   

   It runs scripts/scheduled-ingest.sh every 15 minutes as a low-priority background job. On Linux, run the same script from a systemd user timer or cron; the script is OS-agnostic. (On Windows, the hook needs a PowerShell equivalent of session-end-hook.sh, not yet included; until then, schedule memex ingest-claude-code with Task Scheduler.)

Running always-on (macOS)

memex serve (claude.ai capture) and memex serve-remote (the claude.ai connector) are long-lived servers. To have them start with your Mac and stay up (restarting if they crash), install them as launchd agents. Idle cost is low: neither loads the embedding model until there is real work.

cd /path/to/memex
for svc in serve serve-remote ingest-claude-code; do
  sed "s|__REPO__|$(pwd)|g" "scripts/com.memex.$svc.plist.template" \
    > ~/Library/LaunchAgents/com.memex.$svc.plist
  launchctl load ~/Library/LaunchAgents/com.memex.$svc.plist
done
launchctl list | grep memex      # all three at status 0

serve-remote needs the Tailscale Funnel up on the same port (tailscale funnel --bg 8377); the Funnel persists across reboots. To stop a service: launchctl unload ~/Library/LaunchAgents/com.memex.<name>.plist. (On Linux, run the daemon scripts from systemd user units instead; on Windows, use scripts/install-autostart.ps1 for the capture server.)

Auto-summaries (Phase 3, optional)

When search_chats returns a chat that does not have a summary yet, Memex can generate one on-the-fly using Claude Haiku. The summary is persisted, so the next search of the same chat hits cache and does not pay the API again. This way you only pay for chats you actually look at, not for the whole corpus.

Opt-in. Off by default. Cap: at most 3 summaries generated per search_chats call (parallel), to bound latency and per-query cost.

1. Install the extra (adds the anthropic SDK):

   uv sync --extra summaries
   

2. In your .env (or as env vars), set:

   ANTHROPIC_API_KEY=sk-ant-...
   MEMEX_SUMMARY_ENABLED=true
   

3. Use search_chats from Claude Code as usual. The first time a chat appears in results without a cached summary, Memex generates one. Subsequent searches return the cached summary instantly.

Configurable via env: MEMEX_SUMMARY_MODEL (default claude-haiku-4-5-20251001), MEMEX_SUMMARY_MAX_TOKENS (default 200). If the API fails for a particular chat (no key, rate limit, network), that one result comes back without a summary, the search itself never aborts, and the warning is logged.

Cost model: bulk ingest of an export with 74 chats costs $0 (no summaries are generated at ingest time). Each unique chat you actually open in a search costs roughly $0.01 (Haiku, ~5-10k input + ~200 output tokens). $5 of API credits comfortably covers months of use.

Repo associations (Phase 3)

Memex can associate each chat with the local code repos it touches, and boost those chats when search_chats is invoked from inside a repo. So when you ask Claude Code something like "remember the auth refactor we discussed?", chats that touched the current repo rank higher than unrelated chats with the same keywords.

How it works:

1. Register a repo:

   uv run memex repos add /path/to/your/repo
   

   Memex reads .git/config (origin URL), and pyproject.toml / package.json / Cargo.toml (package name). The canonical key prefers the git remote (stable across clones) over the path.

2. Run a one-time scan over chats you already ingested:

   uv run memex repos scan
   

   Each chat gets matched against every registered repo. The matcher uses 4 signals (highest wins): remote URL literal in chat text (confidence 1.0), absolute path literal (0.9), manifest name word-bounded (0.8), display name word-bounded (0.5). Anything below 0.5 is dropped. New chats ingested after registering the repo are auto-scanned at ingest time.

3. From Claude Code, search with the repo argument:

   search_chats(query="auth refactor", repo="d:/dionisio/memex")
   

   Or pass the git remote URL, or the canonical key from memex repos list. Chats associated to the repo get their distance reduced by 0.3 * confidence. Chats outside the repo still appear lower down (it is a boost, not a filter).

CLI helpers:

uv run memex repos list           # show registered repos
uv run memex repos remove <key>   # unregister + cascade-remove associations
uv run memex tag <chat-uuid> <repo-key>    # manual override (sticky vs auto-scan)
uv run memex untag <chat-uuid> <repo-key>  # remove an association

A manual tag survives subsequent auto-scans: once you tag a chat by hand, the matcher will not overwrite it.

Proactive context injection (SessionStart hook)

Claude Code supports hooks that run shell commands at session boundaries. Memex ships memex session-context, designed to be wired into the SessionStart hook so every new Claude Code session in a registered repo starts with a Markdown blob listing the most relevant past chats. No query needed.

To enable it, add to your .claude/settings.json (project-local) or ~/.claude/settings.json (user-global):

{
  "hooks": {
    "SessionStart": [{
      "command": "uv run memex session-context"
    }]
  }
}

The command auto-detects the active repo by walking up from cwd until it finds .git. If the repo is registered in Memex and has associated chats, it prints a short Markdown blob with the top N (default 5) chats, ordered by manual first, then auto by confidence. If anything fails (no .git, repo not registered, no associations), it prints nothing on stdout (only diagnostics on stderr), so the hook is a silent no-op in that case.

You can also run the command by hand to debug: uv run memex session-context [--repo <path>] [--limit N].

Live capture (Phase 2)

So that new Claude.ai chats land in Memex without asking for a manual export:

1. Start the local HTTP server in a terminal:

   uv run memex serve
   

   Listens on 127.0.0.1:5777 by default. Keep it running while you browse claude.ai. On start it prints an access token; the extension must send it (the loopback Origin check alone does not authenticate other local processes). Reprint it any time with uv run memex token.

2. Load the Chrome extension:

   • Soon: install from the Chrome Web Store (in review for the alpha).
   • For now (unpacked load):
     • Open chrome://extensions/
     • Enable Developer mode
     • Load unpacked → pick chrome-extension/ from the repo
     • Click the Memex icon and confirm the "Server" chip says responding (green).

3. Pair the extension with the access token. Run uv run memex token, copy the value, paste it into the extension popup's token field, and click Save token. This is a one-time step per machine (the token lives user-only next to your DB). Without it, ingest requests are rejected with 401.

4. Use claude.ai normally. Every chat you open or create is ingested automatically. Verify with memex stats or by calling search_chats from Claude Code.

Details in chrome-extension/README.md.

Autostart (Windows + Linux)

So you do not have to run memex serve by hand every time you log in:

memex install-service          # default action: install
memex install-service status   # check current state
memex install-service uninstall

The CLI dispatches to the right installer for your OS:

• Windows: registers a Scheduled Task (MemexServe) that runs uv run memex serve at logon. No admin required, no console window, survives VS Code close. Logs at %LOCALAPPDATA%\Memex\serve.log. Auto-restarts up to 3 times if the wrapper dies.
• Linux: writes a systemd user unit at ~/.config/systemd/user/memex-serve.service, enables it, and starts it now. Logs at ~/.local/state/memex/serve.log. To keep it running across logout: loginctl enable-linger $USER. Status: systemctl --user status memex-serve.
• macOS: the command prints manual start instructions; the supported path is the launchd agents from Running always-on (macOS), which cover serve, serve-remote, and the scheduled ingest in one go.

Making Claude use Memex proactively

By default, LLMs are conservative with tools: they prefer to ask before invoking anything. If you say "remember we talked about X?", Claude tends to answer "I don't recall" instead of searching.

The docstrings of the 3 tools already include "USE PROACTIVELY" instructions, but you can reinforce it by adding this snippet to your CLAUDE.md (global at ~/.claude/CLAUDE.md for every session, or local at <project>/CLAUDE.md for a specific one):

## Memex: persistent memory of Claude.ai chats

There is an MCP server `memex` with 3 tools: `search_chats`, `get_chat`, `list_recent_chats`.
They index ALL of the user's Claude.ai history, reachable via hybrid search
(semantic + lexical FTS5).

**Operational rule:** before answering "I have no record", "I don't remember", "this is
the first time I hear about this", or anything equivalent, call `mcp__memex__search_chats`
with the relevant query. Claude Code's native memory starts clean every session; Memex
is the only path into the user's real history.

Typical triggers: "remember when...", "did I tell you about...", "we already talked about...",
"the other day we discussed...", or any reference to a project / person / decision that might
live in history.

Roadmap

See ROADMAP.md for phases, close criteria, and current status (in Spanish, it is the internal journal).

Devlog

See DEVLOG.md for the log of decisions, blockers, and progress (in Spanish).

Inspiration and references

• Official feature request: anthropics/claude-code#12858
• Claude Historian, claude-conversation-extractor: MCP servers for Claude Code / Desktop history. Reference for tool structure.
• claude-conversation-export: Claude.ai exporter using the same capture strategy. Useful as backfill.
• Spin-off of the SyncChat project.

License

MIT.