| name | memroos-save |
| description | Persist durable agent work product to MemroOS MCP knowledge base with verification, metadata, and source-drift support. Use when producing research, competitive analysis, market positioning, comparison posts, benchmarks, RCA documents, or any "save/document/archive" request. Auto-validates the write succeeded. |
| version | 1.1.0 |
| source_repo | https://github.com/lac5q/memroos/blob/main/.agents/skills/memroos-save/SKILL.md |
Memroos Save — Durable Knowledge Persistence
When to Use This Skill
Use this skill whenever any of these are true:
- The user asked you to "save," "document," "archive," "file," or "store" something
- You produced research, a competitive analysis, a benchmark, or a market-positioning doc
- You wrote a root-cause analysis (RCA) or post-mortem
- You generated a comparison table (X vs Y, A vs B, etc.)
- You produced content that took 5+ tool calls to assemble
- A previous turn delivered analysis as chat output and the user might want it durable
Implicit triggers (the rule fires without explicit ask):
- You produced a
## Comparison, ## Benchmark, ## RCA, ## Analysis, or ## Recommendations section
- You cited external URLs or named sources in your output
- You derived findings from prior research (a follow-up to an existing artifact)
The rule (canonical, in agents/AGENTS_TEMPLATE.md of the MemroOS repo):
When asked to save/document/archive, write to MemroOS first via mcp_memroos_knowledge_write.
This skill exists because agents used to read MemroOS extensively but skip the write step (a "research without persist" failure mode — see RCA in content/research/memroos-persist-failure-rca-2026-07-05.md in the MemroOS repo).
Steps
-
Sanity-check availability. If mcp_memroos_knowledge_write is NOT in the current session's tool list:
- Register MemroOS MCP on the current profile (run
bash $HOME/github/memroos/scripts/install-agent-integrations.sh), OR
- Fall back to direct write to
$MEMROOS_ROOT/content/<topic-slug>.md + git commit && git push, then load the canonical rule + alert the user that the MCP path was bypassed.
-
Choose the path. Default: content/<topic-slug>/<doc-slug>-YYYY-MM-DD.md. Match any existing content/<topic>/ directory if one already exists for the topic. Use lowercase-kebab-case for both the topic and doc slug.
-
Build the markdown. Include this frontmatter:
---
title: "..."
description: "..."
publishedAt: "YYYY-MM-DD"
tags: [...]
keywords: [...]
author: "<agent-name>"
source_session: "<session-id-if-known>"
model: "<model-id>"
sources:
- "https://..."
- "label:identifier"
derived_from:
- "<path-to-prior-artifact>"
regen_prompt: "<prompt that could regenerate this doc>"
---
These metadata fields close the Artyfacts gap — see "Metadata parity with Artyfacts" below.
-
Write it.
mcp_memroos_knowledge_write(
path="content/<topic-slug>/<doc-slug>-YYYY-MM-DD.md",
content=<full_markdown_with_frontmatter>,
auto_commit=True,
commit_message="<type>: <short title>"
)
-
Verify. Read it back:
mcp_memroos_knowledge_read(path="content/<topic-slug>/...")
Confirm the round-trip matched what you intended.
-
Confirm to the user. Quote the path back to the user so they can verify. Do NOT just say "saved" — give them the exact path.
Metadata Parity With Artyfacts
The comparison doc at content/research/memroos-vs-artyfacts.md identified four gaps vs. Artyfacts. This skill requires metadata that closes three of them via frontmatter (the fourth — sharing primitives — is a future MemroOS feature):
| Artyfacts feature | MemroOS equivalent (this skill) |
|---|
| Model attribution per section | model: in frontmatter (per-document for now) |
| Sources / citations | sources: array in frontmatter |
| Regeneration prompt | regen_prompt: in frontmatter |
| Derived-from links | derived_from: array in frontmatter |
| Sharing / public links | Not yet — see Artyfacts gap roadmap |
When you write a doc, fill in what you can:
model — the LLM you used. This is how future readers know who wrote it.sources — every URL you cited, every external doc you referenced, every Discord message you read. The detector at $MEMROOS_ROOT/scripts/source-drift-detector.py will warn if any of these go 404 or change substantially.regen_prompt — a single sentence describing how to regenerate the doc. Useful for "this comparison is now stale, can you refresh it?"derived_from — paths to prior artifacts this one builds on. Forms a derivation chain.
Don't fill these in for trivial one-liners. Do fill them in for research, comparisons, and RCAs.
Common Failure Modes
- "I forgot to call write." This is the most common. Treat
mcp_memroos_knowledge_write as a default — call it for any analysis longer than 3 paragraphs.
- "MCP not in my tool list." Different sessions load different MCP servers. Run
bash $HOME/github/memroos/scripts/install-agent-integrations.sh to register MCP on the current profile.
- "I wrote to Discord/Slack/chat only." Chat is ephemeral. MemroOS is durable. They are not substitutes.
- "The user didn't say 'save.'" If the user asked for a comparison, RCA, benchmark, or research, the implicit ask is "and remember this." Always persist.
- "I called write but didn't include metadata." Even if the user doesn't ask for it, fill in
model, sources, and derived_from. Future-you will thank present-you.
Related Skills
memroos-operations — general MemroOS platform operationsknowledge-base-manager — broader KB management (umbrella)memroos-filing — canonical filing patterns
Sources
- Thread #memroos Discord, July 5 2026: "Microsoft IQ vs MemroOS comparison never persisted to MemroOS"
- RCA:
content/research/memroos-persist-failure-rca-2026-07-05.md (in MemroOS repo)
- Comparison:
content/research/memroos-vs-artyfacts.md (in MemroOS repo)
- Hardening round 2:
content/research/memroos-hardening-july-2026.md (in MemroOS repo)
- Canonical template:
agents/AGENTS_TEMPLATE.md (in MemroOS repo)
- Installer:
scripts/install-agent-integrations.sh (in MemroOS repo)
- Source drift detector:
$MEMROOS_ROOT/scripts/source-drift-detector.py
