enterprise-kb-routing

name	enterprise-kb-routing
description	Use on every run inside Mempedia so the agent behaves like an enterprise knowledge-base operator: search evidence first, update the correct memory layer, and treat skills as reusable operating procedures.
metadata	{"category":"mempedia","priority":"high","always_include":true,"tags":["mempedia","enterprise-kb","routing"]}

Enterprise KB Routing

Goal

Keep the agent operating as part of an enterprise knowledge system rather than as an isolated assistant.

Core Behavior

1. Search before asserting when repository or memory evidence may exist.
2. Prefer the five top-level tools only: read, search, edit, bash, web.
3. Treat read/search/edit as workspace-only tools. All Mempedia Layer 1/2/3/4 operations must go through bash by calling the mempedia CLI.
4. Prefer the narrowest valid memory layer instead of over-promoting information.
5. Let the independent post-turn memory agent classify the full turn across all four layers.

CLI Pattern

• Run from the project root; bash already starts there.
• Resolve the binary before issuing actions:

BIN="${MEMPEDIA_BINARY_PATH:-./target/debug/mempedia}"
[[ -x "$BIN" ]] || BIN=./target/release/mempedia

• For simple actions with no text content (search, list, read), use a single-line heredoc:

cat <<'JSON' | "$BIN" --project "$PWD" --stdin
{"action":"list_skills"}
JSON

• CRITICAL — Writing rich content (Chinese, quotes, newlines, web-sourced text): Heredoc JSON is INVALID when a string value contains literal newlines or unescaped double-quotes. Any content that includes multi-line markdown, non-ASCII text, or external webpage content MUST use the two-step pattern below.

  Step 1: write the markdown to a temp file via a shell heredoc (no Python triple-quotes = no risk of """ in content breaking the script). Step 2: Python reads the temp file and sends it through json.dumps() so all escaping is handled automatically.

  Use the unique delimiters ____MEMPEDIA_MD____ and ____MEMPEDIA_PY____ — content will almost never contain these exact strings.

BIN="${MEMPEDIA_BINARY_PATH:-./target/debug/mempedia}"
[[ -x "$BIN" ]] || BIN=./target/release/mempedia

# Step 1: write markdown to a temp file
MDTMP=$(mktemp)
cat > "$MDTMP" <<'____MEMPEDIA_MD____'
---
node_id: "example_node"
title: "Example Title"
---

# Example Title

Full markdown body here — can contain Chinese, quotes, newlines, code blocks, anything.
____MEMPEDIA_MD____

# Step 2: Python reads the file and sends via CLI with proper JSON escaping
MDTMP="$MDTMP" python3 - <<'____MEMPEDIA_PY____'
import json, subprocess, os

BIN = os.environ.get('MEMPEDIA_BINARY_PATH', './target/debug/mempedia')
if not os.path.isfile(BIN):
    BIN = './target/release/mempedia'

with open(os.environ['MDTMP']) as f:
    markdown = f.read()

payload = json.dumps({
    "action": "agent_upsert_markdown",
    "node_id": "example_node",
    "markdown": markdown,
    "importance": 1.9,
    "agent_id": "agent-main",
    "reason": "User-requested knowledge ingestion",
    "source": "web",
})
result = subprocess.run(
    [BIN, '--project', os.getcwd(), '--stdin'],
    input=payload, capture_output=True, text=True
)
print(result.stdout or result.stderr)
____MEMPEDIA_PY____
rm -f "$MDTMP"

• NEVER use Python triple-quoted strings (""") for markdown content — web-sourced content or Chinese text may contain """ which terminates the string prematurely.
• NEVER embed a multi-line string value inside <<'JSON'...JSON — JSON does not allow literal newlines in string values.
• The agent_upsert_markdown action accepts a full markdown string (YAML frontmatter + body) and correctly derives summary, title, and structured fields from it.

When To Escalate To Memory

• Stable project facts: Layer 1 core knowledge.
• Short-lived chronology: Layer 2 episodic memory.
• Persistent user constraints: Layer 3 preferences.
• Reusable procedures: Layer 4 skills.

Avoid

• Treating skills as answer content.
• Treating a skill name as a tool name.
• Skipping repository evidence when the question is project-specific.