From memex
Guides creation of atomic memory cards with metadata enrichment, candidate retrieval, and deliberate linking for structured knowledge capture.
How this skill is triggered — by the user, by Claude, or both
Slash command
/memex:memex-agentic-memoryThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
An A-MEM-inspired workflow that uses existing memex primitives to produce
An A-MEM-inspired workflow that uses existing memex primitives to produce higher-quality memory cards through structured observation, enrichment, retrieval, and deliberate linking.
This skill is experimental and requires opt-in.
Before proceeding, locate .memexrc using the same precedence as core memex:
$MEMEX_HOME/.memexrc (if MEMEX_HOME env var is set).memexrc~/.memex/.memexrc (fallback)Check that experimental.agenticMemory is exactly true.
If the flag is missing, false, or not exactly true, STOP. Use the standard
memex-retro skill instead. Do not proceed with the agentic workflow.
Three equivalent interfaces — use whichever your environment supports:
| CLI (memex in PATH) | Plugin CLI fallback (Claude Code) | MCP tool (VSCode / Cursor) |
|---|---|---|
memex search <q> | node ~/.claude/plugins/cache/cc-plugins/memex/*/dist/cli.js search <q> | memex_search with query arg |
memex read <slug> | node ~/.claude/plugins/cache/cc-plugins/memex/*/dist/cli.js read <slug> | memex_read with slug arg |
memex write <slug> | node ~/.claude/plugins/cache/cc-plugins/memex/*/dist/cli.js write <slug> | memex_write with slug arg and body |
Resolution order: Try memex in PATH first. If not found, use the plugin
fallback or MCP tools.
Identify whether there is a memory-worthy insight from the completed work.
Skip if the content is:
If nothing is worth remembering, stop here.
For each insight, draft one card:
yaml-array-roundtrip-bug)Generate candidate metadata as simple string fields:
---
title: <title>
created: <YYYY-MM-DD>
source: <client>
category: <optional>
context: <one sentence explaining domain and purpose>
keywords: <3-8 salient terms, comma-separated>
tags: <broad categories, comma-separated>
---
Important: Store context, keywords, and tags as comma-separated
strings, not YAML arrays. This avoids frontmatter serialization issues with
the current stringifyFrontmatter implementation.
Before writing, search for related existing cards.
Try semantic search first:
memex search "<topic query>" --semantic --compact --limit 8
If semantic search fails or is unavailable, fall back to keyword search:
memex search "<topic query>" --compact --limit 8
Read the top candidates that look relevant:
memex read <slug>
Limits to avoid runaway retrieval:
Choose exactly one primary action per insight:
| Action | When |
|---|---|
| create | No existing card covers the insight. Embed [[wikilinks]] in the new card pointing to related candidates when meaningfully related. |
| update | Existing card covers the same insight but lacks new detail |
| skip | Insight is duplicate, too obvious, or not durable |
Rules:
Before writing, produce a preview of planned changes:
Planned memory changes:
- create: <slug> (<title>)
- update: <slug> because <reason why update is better than new card>
- links: [[a]], [[b]] because <relationship explanation>
- metadata: context/keywords/tags draft
For updates to existing cards, the agent MUST explain why updating is preferable to creating a new card.
Use existing memex write paths:
memex write <slug> << 'EOF'
---
title: <title>
created: <YYYY-MM-DD>
source: <client>
category: <optional>
context: <one sentence>
keywords: <comma-separated terms>
tags: <comma-separated categories>
---
<One atomic insight in your own words.>
This relates to [[existing-card]] because <explicit relationship explanation>.
EOF
On update, copy ALL existing frontmatter fields from the current card (title, created, source, category, context, keywords, tags, and any custom fields). Only modify the specific fields previewed in Step 6. Append new information to the body.
After writing:
memex read <slug>
Confirm:
[[slug]] format)[[links]] must appear in sentences explaining the relationship.
memex-retro instead.npx claudepluginhub iamtouchskyer/memex --plugin memexReflect on completed tasks and save atomic insights to a Zettelkasten memory system. Invoke after code changes, debugging, or non-trivial problem solving.
Captures user thoughts, ideas, knowledge, questions, or sources as TwinMind cards with type classification, duplicate detection, creation, auto-linking, and index updates. Use for memorable knowledge shares.
Manages persistent semantic memory across sessions: store/retrieve knowledge/TODOs/issues, hybrid semantic search, hierarchy/tags organization, and maintenance tools.