feat(sources): add mengram plugin

[alibaizhanov]

Type of Change

• 🔌 New plugin submission

Description

Summary: Adds mengram to sources.yaml so it auto-syncs into plugins/productivity/mengram.

Motivation: Mengram is persistent memory for Claude Code: hosted backend, hybrid retrieval (vector + BM25 + RRF), Ebbinghaus temporal decay, and importance weighting. The plugin under integrations/claude-code-plugin wires SessionStart + Stop hooks, a memory skill, and an MCP server into Claude Code so context survives /clear, machine switches, and team handoffs.

Related Issues: N/A

Plugin Details

Plugin Name: mengram
Category: productivity
Version: 0.1.0

Components Included:

• Skills (skills/memory/SKILL.md)
• Hooks (SessionStart, Stop)
• MCP server (mengram via Streamable HTTP at mengram.io/mcp)

Dependencies: Hosted Mengram API (free tier at https://mengram.io, 40 adds + 200 searches/mo). API key via MENGRAM_API_KEY env var.

Checklist

• Plugin has valid .claude-plugin/plugin.json (passes claude plugin validate)
• README.md is comprehensive (in integrations/claude-code-plugin/README.md)
• LICENSE file is included (Apache-2.0, repo root)
• No hardcoded secrets, API keys, or credentials
• Hooks use ${CLAUDE_PLUGIN_ROOT} for portable paths
• All JSON files valid

Testing Evidence

Test Environment: macOS, Claude Code 2.1.175

Test Commands Run:

claude plugin marketplace add alibaizhanov/mengram
claude plugin install mengram@mengram
claude plugin details mengram

Plugin installed successfully, status enabled, hooks fire on SessionStart and Stop, MCP server connects to mengram.io/mcp.

Breaking Changes

• No

Security Considerations

• No hardcoded secrets — API key read from MENGRAM_API_KEY env
• All network calls documented in README (HTTPS to mengram.io/v1/*)
• Minimal permissions

Also submitted to

• Official Anthropic claude-plugins-community marketplace (2 days ago, pending review)
• claudemarketplaces.com (today via hi@ email)

[jeremylongshore]

Thanks for submitting Mengram — the plugin architecture is genuinely solid. The hook scripts are clean, the README is thorough about privacy and data flow, and the SKILL.md body (when-to-recall / when-NOT-to-recall / when-to-capture sections) is one of the better memory-skill write-ups we've seen.

One blocker before this can merge: the marketplace-tier validator requires 8 frontmatter fields in every SKILL.md, and skills/memory/SKILL.md currently only declares name and description. The six missing fields are:

allowed-tools: Read          # list the MCP tools the skill actually invokes, e.g. mcp__mengram__search
version: 0.1.0               # match plugin.json
author: Ali Baizhanov <ali@mengram.io>
license: Apache-2.0
compatibility: Designed for Claude Code
tags: [memory, mcp, productivity]

One note worth surfacing for users in the README or SKILL.md: the Stop hook sends the last ~8 KB of the conversation transcript to mengram.io/v1/add_text on every turn. That's the whole point of the product and you document it well under "Privacy" — just worth making equally prominent in the skill's frontmatter description or a callout so installers see it before the first hook fires.

Fix those six fields and this is ready to re-review. The content itself is merge-quality.

One item under separate review (not a fields blocker): the Stop hook (save-conversation.sh) sends the last ~8KB of every conversation transcript to mengram.io/v1/add_text. It's disclosed in your README and no-ops without an API key — that's honest design. But shipping conversation content to a hosted third party is a marketplace trust consideration we weigh on its own merits before listing. Flagging it transparently; no action needed from you on that point yet.

---

Audited against the marketplace rubric (frontmatter completeness, license accuracy, tool-safety, real-vs-hollow content). Holding open as request-changes — happy to merge once the above lands.

• Jeremy Longshore
  intentsolutions.io