🗄️ Anthropic이 잊어도 Memex는 기억합니다

Your AI session history
survives Anthropic's silent migrations.

Claude Code rewrites its session storage every few months without announcement and auto-updates silently delete the old files (#41591 · #54907 · #48782). Memex reads both the legacy ~/.claude/transcripts/ and modern ~/.claude/projects/ paths, indexes them into a local Qdrant you own, and gives you a snapshot button.

Vannevar Bush imagined the original Memex in 1945 — a personal knowledge machine built on associative trails, not search boxes. Eighty years later this is its desktop reincarnation — wrapped in a non-chatbot UI powered by six Qdrant primitives.

View on GitHub ↗ Install ↓ See predict in action →

Works as an MCP server — connect any AI agent (Claude Code, Codex, Cursor) to your local memex

# one-time wiring — any MCP-aware AI agent can now call Memex tools
claude mcp add memex /path/to/memex/src-tauri/target/release/memex mcp

# or let Memex print the exact line for your machine
memex install-mcp --run

9 tools exposed: find_similar_sessions · find_similar_error · predict_next_action · mix_similar_sessions · get_session_summary · get_session_turn · list_recent_sessions · analyze_corpus_topology · snapshot_export. All stay 100 % local — no telemetry, no LLM call inside the server.

🪟 Time Machine stack

🌌 Topology galaxy

⏯ Replay engine

Why this isn't a chatbot

Qdrant Vector Space Day 2026's prompt is unusually direct:

"Think Outside the Bot." Forget the classical RAG chatbot. Reimagine vector search beyond conversational interfaces — multi-modal apps, intelligent recommendations, advanced vector search. — Qdrant VSD 2026 hackathon prompt

The obvious RAG-chatbot version of this	What Memex does instead
A text box asking "what session am I looking for?"	A 3D Time Machine card deck showing every past session — ↑↓/wheel to time-travel.
Embed + retrieve + LLM-summarize a session.	Replay the session turn-by-turn with Bash terminals, Edit diffs, Read snippets — exactly as you saw them live.
"Have I seen this error before?" → RAG → LLM → text answer.	A banner slides in with the past session whose error named-vector neighborhood matches — zero LLM calls.
"What's the shape of my work?" → LLM writes a paragraph.	3D topology galaxy from Qdrant's Distance Matrix API, with auto-labeled clusters and cross-project bridge edges.
"What should I do next?" → LLM completion + tool-use.	Predict next-action — embed the active session, find similar past sessions in Qdrant, surface what past-you did from a comparable position. No LLM.

6Qdrant primitives, six surfaces

Each surface in Memex maps to a different Qdrant primitive — together they cover named vectors → matrix sampling → discovery → payload filtering → snapshots → recommendation. None of these are the "embed-and-summarize" loop of classical RAG.

🪟

Time Machine stack

payload-only scroll · no vectors

Boot the app and every past session appears as a 3D layered card deck. Arrow keys time-travel. No search box. Ever.

🌌

Topology galaxy

Distance Matrix API · search_matrix_pairs

WebGL scene of your corpus. Force-directed cluster auto-labels ("code + shell · Bash×1350 Edit×1032"). Yellow bridges = ideas that jumped between projects.

🧪

Mix & Match

Discovery API · ContextInputPair

Drop sessions as positives + negatives. Qdrant returns sessions near the positives and far from the negatives. Recommendation, not retrieval.

🔔

Proactive recall

error named-vector + has_errors filter

While you work in another Claude Code session, Memex watches for fresh tool errors. Match a past resolution? A banner slides in with the link.

NEW

🔮

Predict next-action

content vector + neighbor pivot walk

Embed your last few turns, find K similar past sessions, locate the pivot turn, walk forward — surface what past-you did next. Ranked by frequency × similarity.

⏯

Replay engine

payload source_path · on-demand re-parse

Turn-by-turn animation of any past session with Bash terminals, Edit -/+ diffs, Read snippets. ⏮ ⏯ ⏭ controls, 1×/2×/4×/8×.

🔍

Lens slider

5 named vectors per point · parallel query

The "advanced vector search" axis — intentionally last. Five named vectors (content, tool, path, error, code), weighted by sliders. Per-vector contribution chips on every hit.

📦

Snapshot

Qdrant HTTP snapshot API

One portable file holding your entire indexed memory. Export, archive, share, restore on another machine.

🔮 What past-you did next

The killer surface for a hackathon themed "intelligent recommendations". Click any session — Memex queries the corpus for similar past sessions, finds the conversational pivot in each one, walks horizon turns forward, and aggregates the tool calls.

Zero LLM — pure vector neighbors + tool-call frequency.
Cards rank by frequency × similarity.
Each pick links back to the exact source-turn-index — click to Replay it.
Mines your own engineering history as ground truth.

🖥 Bash · 67% · sim 65%

cargo build --release

● project-redesign · turn #486 →

✏️ Edit · 33% · sim 64%

src-tauri/Cargo.toml

● project-tool-a · turn #312 →

Architecture

One desktop binary. One local Qdrant. Zero network calls after install. Your ~/.claude/projects/**/*.jsonl never leaves the machine.

~/.claude/projects/ ──[ walkdir + serde_json ]──▶ parser.rs │ ▼ 5 string extracts ┌──────────────────────────┐ │ indexer.rs │ │ fastembed BGE-small │ │ qdrant-client 1.18 │ │ predict_next_actions() │ └────────┬─────────────────┘ │ gRPC / HTTP ▼ ┌──────────────────────────┐ │ local Qdrant 1.18 │ │ 5 named vectors × point │ │ payload-indexed metadata │ └──────────────────────────┘ ↑ Tauri 2 webview ──▶ vanilla JS UI ↑ commands.rs ── thin IPC wrappers

Install

macOS 11+ (Apple Silicon recommended). Tested on macOS 26.5 / arm64. Linux + Windows builds are roadmap items.

Clone & install JS deps

gh repo clone sgwannabe/memex ~/memex
cd ~/memex
npm install

Start Qdrant locally

# Binary (arm64 macOS):
mkdir -p .qdrant && curl -sL \
  https://github.com/qdrant/qdrant/releases/download/v1.18.0/qdrant-aarch64-apple-darwin.tar.gz \
  | tar xz -C .qdrant
./.qdrant/qdrant &

Index `~/.claude/projects`

cargo build --release --manifest-path src-tauri/Cargo.toml
./src-tauri/target/release/memex scan --index
# ~130 MB BGE-small ONNX downloads on first run, cached.

Launch the app

npm run tauri build
open src-tauri/target/release/bundle/macos/Memex.app

Your AI session historysurvives Anthropic's silent migrations.