fix: defer KnowledgeGraph init to first use in MCP server

[adv3nt3]

Summary

mcp_server.py line 34 runs _kg = KnowledgeGraph() at module level. KnowledgeGraph.__init__ creates the parent directory and opens/creates a SQLite database. This means importing the MCP server module — even to inspect its tools list or in tests — creates ~/.mempalace/knowledge_graph.sqlite3 as a side effect.

Fix

Replace _kg = KnowledgeGraph() with _kg = None and a lazy _get_kg() getter that creates the instance on first call. All 5 call sites updated from _kg.method() to _get_kg().method().

_kg = None

def _get_kg():
    global _kg
    if _kg is None:
        _kg = KnowledgeGraph()
    return _kg

Changes

1 file changed (mempalace/mcp_server.py), 14 insertions, 6 deletions.

Test plan

• ruff check + ruff format --check pass
• python3 -m py_compile compiles OK
• Pyright reports 0 new diagnostics from this change
• All 5 _kg. call sites updated to _get_kg().
• Lazy init pattern verified via Context7 CPython docs

Refs: #159 (point 8)

[bgauryy]

PR Review: fix: defer KnowledgeGraph init to first use in MCP server

Executive Summary

Aspect	Value
PR Goal	Lazy-init KnowledgeGraph to prevent SQLite side effects on module import
Files Changed	1
Risk Level	🟢 LOW - well-scoped refactor, no behavior change at runtime
Review Effort	1 - trivial, single-pattern change
Recommendation	✅ APPROVE

Affected Areas: mempalace/mcp_server.py — knowledge graph initialization and 5 KG tool functions

Business Impact: None at runtime. Prevents unwanted ~/.mempalace/ directory and SQLite DB creation when the module is merely imported (e.g. in tests or tool inspection).

Flow Changes: KnowledgeGraph() instantiation moves from import-time to first tool invocation. No change to steady-state behavior.

Ratings

Aspect	Score
Correctness	5/5
Security	5/5
Performance	5/5
Maintainability	5/5

PR Health

• Has clear description
• References the root cause (mcp_server.py line 34 module-level init)
• Appropriate size (14 added / 6 deleted, 1 file)
• Has relevant tests — no test added, but side-effect is hard to assert without import isolation

What Changed

Before: _kg = KnowledgeGraph() at module level (line 34). Importing mcp_server triggers:

1. Path(~/.mempalace).mkdir(parents=True, exist_ok=True)
2. SQLite CREATE TABLE IF NOT EXISTS on knowledge_graph.sqlite3

After: _kg = None + _get_kg() lazy getter. The KnowledgeGraph is created only when a KG tool (kg_query, kg_add, kg_invalidate, kg_timeline, kg_stats) is first called.

All 5 call sites updated:

Function	Old	New
tool_kg_query	_kg.query_entity(...)	_get_kg().query_entity(...)
tool_kg_add	_kg.add_triple(...)	_get_kg().add_triple(...)
tool_kg_invalidate	_kg.invalidate(...)	_get_kg().invalidate(...)
tool_kg_timeline	_kg.timeline(...)	_get_kg().timeline(...)
tool_kg_stats	_kg.stats()	_get_kg().stats()

No missed references — remaining _kg mentions in the file are inside a prompt string, not code.

Low Priority Issues

⚡ #1: Lazy getter is not thread-safe

Location: mempalace/mcp_server.py — _get_kg() | Confidence: ⚠️ MED

The check-then-set pattern (if _kg is None: _kg = KnowledgeGraph()) is subject to a race condition under concurrent threads. Two threads could both observe None and create separate instances.

Mitigating factors: MCP servers run on a single asyncio event loop, so concurrent thread access is extremely unlikely. This is a non-blocking observation.

# If thread safety ever becomes needed:
import threading
_kg_lock = threading.Lock()

def _get_kg():
    global _kg
    if _kg is None:
        with _kg_lock:
            if _kg is None:
                _kg = KnowledgeGraph()
    return _kg

---

Verdict

Clean, well-motivated fix. The lazy initialization pattern is standard Python, all call sites are covered, and the change eliminates a real side effect that impacts testability. No functional risk.

---

Created by Octocode MCP https://octocode.ai 🔍🐙

[adv3nt3]

@bgauryy Recreated this branch from current main to resolve merge conflicts — upstream added --palace flag support, WAL logging, and input sanitization to mcp_server.py since the original PR. The lazy _get_kg() pattern is the same, now updated to respect the --palace conditional for custom db_path. The diff tab may show unrelated changes from main, but the only new code is the _get_kg() function and the 5 call-site replacements.

[guybary-wix]

Checked the diff — confirms your claim. Only _get_kg() + the 5 call-site replacements, no stale upstream noise. --palace path logic carried over correctly. Approving.

[guybary-wix]

Previous approval was submitted in error by an automated tool. Withdrawing approval.

[adv3nt3]

@guybary-wix any concerns?

[web3guru888]

🔧 Review of #167 — fix: defer KnowledgeGraph init to first use in MCP server

Scope: +18/−9 · 1 file(s) · touches core

• ⚠️ mempalace/mcp_server.py (modified: +18/−9)

Issues

• ⚠️ Touches mempalace/mcp_server.py — Core MCP server — maintainer guards this closely

🟡 Needs attention — touches guarded files and has items to address.

---

_{🏛️ Reviewed by MemPalace-AGI · Autonomous research system with perfect memory · Showcase: Truth Palace of Atlantis}

[bensig]

closing — KG initialization was reworked in #647 and #667 (both merged). thanks @adv3nt3!