memex

Memex

Structured memory for Claude Cowork. Pick up where you left off.

Persistent memory across Cowork sessions in pure markdown. The closets-format index matches full-content keyword search on retrieval recall at roughly 1/10th the size, with zero external dependencies. No database, no API keys, no embeddings backend.

---

The Problem

Cowork sessions are stateless. You open a new session and Claude has no idea what you were doing yesterday. You spend the first 10 minutes re-explaining context. Decisions from three sessions ago? Gone.

Cowork Projects gives you persistent files, but files without a system is just a folder full of stuff. Claude doesn't know what to read, what's stale, or what matters right now.

The Solution

Memex converts your workspace into a connected knowledge system with persistent memory, tiered context loading, full [[wikilink]] navigation, and cross-workspace federation. All in pure markdown with zero runtime dependencies.

• Wikilinked knowledge base. Every file reference becomes a [[wikilink]]. Your workspace builds into a connected graph over time. Open it in Obsidian to see how everything relates visually.
• Two-tier index. A _MANIFEST.md plus per-hub _CLOSETS.md files (and memory/_CLOSETS.md for Tier 1) mean Claude knows what every file contains without opening any of them. Field-level retrieval (subjects, people, claims, decisions, dates, status) on questions about specific subjects, not just topics.
• Typed-edge graph. Optional YAML frontmatter (supersedes, blocks, people, projects) builds a typed knowledge graph. Zero LLM calls; pure regex.
• Cross-workspace federation. Register multiple workspaces (nonprofit, personal, work) in a global registry and search across all of them with /memex:cross-search. Privacy-first: opt-in per source.
• Cross-hub search within a workspace. /memex:search greps the manifest plus every closets file, grouped by folder.
• Standalone consolidation cycle. /memex:consolidate runs dedup, decisions contradictions, orphan check, and decisions compression independently from session-end, so a session timeout doesn't compound drift.
• Convention over configuration. Drop files in standard locations and they just work. No config tables to maintain.
• Zero dependencies. Markdown files plus stdlib Python. No database server, no API keys, no embeddings backend, no cloud.

Retrieval benchmark

Reproducible on LongMemEval-S (Wu et al., ICLR 2025). The closets-format index hits 90.1% Recall@5 with the closets:emax strategy, within 0.5pp of content:bm25 (the upper bound that indexes the entire raw session text) at roughly 1/10th the size. 500 questions, free, deterministic, runs in 3-5 minutes:

python benchmarks/longmemeval/run_bench.py --strategies closets:emax

See benchmarks/longmemeval/ for the harness, ablations, per-category breakdown, and the full retrieval-vs-QA-accuracy context.

---

Install

Requires Claude Cowork. (Can also be adapted for Claude Code.)

1. Add the Memex marketplace. Either method works:

   • Cowork UI: Click Customize > Browse plugins > "+" to add marketplace. Enter Skyfox-io/Memex and click Sync.
   • Slash command: Type /plugin marketplace add Skyfox-io/Memex in any session.

   (Skyfox-io/Memex is GitHub shorthand for github.com/Skyfox-io/Memex.)

2. Install the plugin when prompted

3. Run /memex:init in your workspace

New workspace? Init asks what you're working on, then scaffolds everything in under 30 seconds.

Existing files? Init scans your workspace, shows you what it found, and wires everything into a manifest without moving files.

Already have a memory system? Init detects existing manifests and operates in compatible mode.

After setup

Start a new Cowork session. Claude briefs you automatically and asks what you're working on.

If you use Obsidian, open the same workspace folder as a vault. All the [[wikilinks]] light up in the graph view.

---

How It Works

Closets, the typed-field index

Each domain has a _CLOSETS.md file (and memory/_CLOSETS.md covers Tier 1) that enumerates every file's distinct subjects, named entities, claims, decisions, and dates. Sessions scan closets to pick which 0-2 files a question actually needs, without opening any of them.

Sample closets entry:

## [[campaign-plan]]
- subjects: spring fundraising, gala, donor outreach
- people: [[Mike]], [[Jasmine]]
- claims: "raised $42K so far"
- decisions: switched from HubSpot to Anchor on 2026-03-18
- dates: 2026-04-15 (gala)
- status: active