Neuro MCP Server

neuromcp

Local-first MCP memory server with closed-loop attribution critic. Oracle-split and distractor-split benchmark numbers both published, with sample-size caveats.

Local-first MCP server with hybrid search, verbatim recall, and crash-resilient session persistence.

npx neuromcp

Benchmarks (v0.18.0)

Oracle split (clean — easy mode)

Oracle-split LongMemEval isolates the correct memory in a small corpus. Every local MCP memory system claims ~99% here. It measures "does the ranker work on clean inputs" — nothing more.

Distractor split (v0.18.0, honest)

Same 30 questions + 1000 random distractor memories drawn from other questions' haystacks. The correct memory now competes against real noise.

Reproduce: npx tsx eval/longmemeval-distractor-runner.ts --limit 5 --distractors 1000

Sample sizes. The 500-distractor row is n=30 (Wilson 95% CI for 28/30 ≈ 78-99% R@5). The 1000-distractor row is n=5 — preliminary, Wilson 95% CI [57%, 100%]. The 1000-distractor n=30 run takes ~36 min on a single Ollama instance; cached-distractor batching is v0.19.0 work. Treat 500-distractor numbers as defensible, 1000-distractor as directionally positive but underpowered.

Head-to-head comparison is explicit v0.19.0 work. Hindsight (local OSS MCP, ~94.6% LongMemEval claimed) and Mem0/Zep publish their own numbers on their own harnesses. Until we run all of them against the same corpus + embedder, calling any local MCP server "state of the art" is marketing, not measurement. neuromcp publishes its numbers with sample-size caveats so you can judge direction; don't read absolute superiority into them yet.

Hybrid ranker (BM25 + vector + attention + graph + usefulness prior) keeps R@5 = 100% at 1000:1 distractor:target ratio on the observed sample. MRR drops to 74% because the correct memory is sometimes not rank-1 but always rank ≤ 5 in what we saw. Earlier v0.18.0 numbers (R@5 23%) were from a test FakeEmbedder — fixed in v0.18.1.

What this benchmark does NOT prove: end-to-end answer correctness, long-horizon multi-session reasoning, or superiority over commercial cloud systems (Mem0, Zep) on their own benchmarks. Those comparisons need their numbers on the same distractor split, which hasn't been published.

Why

AI agents forget everything between sessions. Existing solutions either store flat key-value pairs (useless for real knowledge) or require cloud infrastructure and API keys.

neuromcp gives you two layers of memory:

1. MCP Server — hybrid search (vector + full-text + graph), verbatim recall, memory governance, automatic consolidation, all in a single SQLite file
2. Wiki Knowledge Base — compiled Markdown knowledge that survives crashes, compounds over sessions, and gives your agent project-aware context at every startup

Inspired by Karpathy's LLM Wiki, Mastra's Observational Memory, and Zep's temporal knowledge graphs — but simpler than all of them. No vector DB, no embeddings pipeline, no cloud. Just Markdown files + Git + hooks.

Architecture

~/.neuromcp/
├── memory.db               ← SQLite: hybrid search, MCP tools
├── wiki/                   ← Compiled knowledge (git-tracked)
│   ├── index.md            ← Routekaart — LLM reads this FIRST
│   ├── schema.md           ← Operating rules for the LLM
│   ├── log.md              ← Append-only changelog
│   ├── people/             ← User profiles, preferences
│   ├── projects/           ← Project knowledge (stack, auth, URLs)
│   ├── systems/            ← Infrastructure (tools, MCP servers)
│   ├── patterns/           ← Reusable patterns (error fixes, routing)
│   ├── decisions/          ← Architecture decisions with context
│   └── skills/             ← Repeatable procedures
└── raw/sessions/           ← Raw session logs (auto-generated)

How the wiki works

Self-healing consolidation pipeline (v0.15.0+)

Every ~4h the launchd agent runs run-consolidation.sh, which orchestrates four steps end-to-end:

1. consolidate-sessions.py — batches raw sessions per project, asks Claude for a factual summary, and fact-checks it against the raw sources. When the auditor flags specific unsupported claims the consolidator now auto-strips those lines and re-audits once — so one speculative sentence no longer kills a whole batch.
2. rescue-rejected.py — any batch that still fails is parsed, the unsupported claims are removed, and the cleaned summary is appended to its wiki page. Pure text surgery, no LLM calls.
3. entity-linker.py — cross-links every page: a bare-word mention of another registered entity (people/, projects/, systems/) is added to the page's related: frontmatter. Makes the wiki act like a graph without a separate graph database.
4. rebuild-index.py — regenerates index.md and per-category -index.md files. Categories over 10 pages are auto-split so the session-start router stays compact as the wiki scales.

The pipeline is idempotent — safe to re-run at any time.

What the LLM knows at session start

Schema (operating rules) → How to maintain the wiki
Index (knowledge map)    → What knowledge exists
User profile             → Who you are, how you work
Project page             → Current project details (auto-detected from cwd)
Last session             → What happened last time

Quick Start

1. Start the MCP server

npx neuromcp

2. Initialize the wiki + hooks (required for closed-loop attribution)

npx neuromcp-init-wiki

This creates the wiki structure, installs hooks (Claude Code) and rules (other editors), and configures everything automatically. Without this step, npx neuromcp still runs as a plain MCP server with 42 tools, but the critic hook that closes the attribution loop is not installed — retrieval works but usefulness scores never accumulate. Safe to run multiple times — won't overwrite existing config.

Editor Compatibility

neuromcp works with any MCP-compatible editor. Two tiers of integration:

Claude Code gets the full experience via native hooks — context injection and persistence happen automatically, even if the LLM forgets.

Other editors get rules files that instruct the LLM to call neuromcp tools at session start/end. This depends on LLM compliance — it works well in practice but is not guaranteed like hooks.

# Auto-detect installed editors
npx neuromcp-init-wiki

# Target a specific editor
npx neuromcp-init-wiki --editor cursor

# Install rules for all supported editors
npx neuromcp-init-wiki --editor all

Supported editors: cursor, windsurf, cline, copilot (VS Code), jetbrains, zed

Recommended: Add Ollama for real semantic search

ollama pull nomic-embed-text

neuromcp auto-detects it. No config needed.

Installation

Claude Code

// ~/.claude.json → mcpServers
{
  "neuromcp": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "neuromcp"]
  }
}

Claude Desktop

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "neuromcp": {
      "command": "npx",
      "args": ["-y", "neuromcp"]
    }
  }
}

Cursor / Windsurf / Cline

Same format — add to your editor's MCP settings.

Per-project isolation

// .mcp.json in project root
{
  "mcpServers": {
    "neuromcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "neuromcp"],
      "env": {
        "NEUROMCP_DB_PATH": ".neuromcp/memory.db",
        "NEUROMCP_NAMESPACE": "my-project"
      }
    }
  }
}

MCP Surface

Core Tools

Verbatim Tools

Resources (13)

Prompts (3)

Wiki Knowledge Base

The wiki is the compiled, human-readable knowledge layer. It replaces the chaos of session logs with structured, interlinked Markdown pages.

Why a wiki instead of more vector search?

Wiki page format

---
title: My Project
type: project
created: 2026-04-06
updated: 2026-04-06
confidence: high
related: [other-project, oauth-setup]
---

# My Project

Description, stack, auth, deployment details...

How to use

The wiki works automatically once hooks are installed. The LLM:

1. Reads index.md at session start to know what knowledge exists
2. Reads specific pages when relevant to the current task
3. Updates pages when learning something new
4. Gets reminded every 8 tool calls if the wiki needs updating

You can also browse and edit the wiki manually — it's just Markdown files.

Auto-consolidation (optional)

Once you accumulate raw session logs, the wiki can be kept fresh automatically. A scheduled job reads unprocessed sessions, groups them per project (by detecting $HOME/projects/<name> paths in the session content), and uses the claude CLI to synthesise a ## [date] entry into the right wiki page.

npx neuromcp-enable-consolidation

What it installs:

• ~/.neuromcp/scripts/consolidate-sessions.py — the worker
• ~/.neuromcp/scripts/run-consolidation.sh — threshold-guarded runner
• macOS: a launchd agent that fires every 4 hours (com.neuromcp.consolidate)
• Linux: prints a cron snippet to add manually

Requirements:

• python3 ≥ 3.8 on PATH
• the claude CLI on PATH

Guards built in:

• Threshold: skip if fewer than 5 unprocessed sessions
• Output is extracted from a fenced markdown block; apology/narration text is rejected
• Ledger (~/.neuromcp/consolidation-ledger.json) makes re-runs idempotent
• Large project backlogs are auto-batched (default 15 sessions per claude call; override with --max-sessions)

Uninstall: npx neuromcp-enable-consolidation --uninstall

Change interval: npx neuromcp-enable-consolidation --interval 7200 (every 2 hours)

Hallucination guard (eval-loop). Every consolidator output goes through a second Haiku audit before the wiki is touched. If any factual claim in the generated summary is not traceable to the raw sessions, the chunk goes to ~/.neuromcp/review-queue/ instead of the wiki. No hallucinated claims leak through.

Atomic facts with temporal supersession. After a summary is approved, it is also distilled into short standalone facts and stored as category='fact' rows with valid_from=today. When a new fact is Jaccard-similar to an existing one in the same project, Haiku decides whether NEW supersedes OLD — if yes, the old row gets superseded_by_id and valid_to set. Retrieval defaults to current facts only (superseded_by_id IS NULL), so outdated conclusions never resurface.

Auto-retrieve + hybrid indexing

Once the wiki has content, make it searchable so the UserPromptSubmit hook can surface relevant pages automatically (no more "LLM must remember to call search"):

npx neuromcp-index-wiki              # index wiki pages into memories_fts + memories_vec
npx neuromcp-index-wiki --rebuild    # wipe wiki entries first, then reindex
npx neuromcp-index-wiki --dry-run    # preview what would change
npx neuromcp-index-wiki --no-embed   # FTS-only mode (no embedding provider needed)

npx neuromcp-backfill-embeddings     # embed any memory still missing a vector

The indexer splits each page on ## section headers and stores every section as a deduplicated memory (source='wiki', category='wiki'). Each section is both written to the FTS5 index and embedded via the configured provider (Ollama → OpenAI → ONNX) so vector search works too.

At prompt time the neuromcp-auto-retrieve.js hook calls neuromcp-query, which runs FTS5 BM25 and sqlite-vec cosine search in parallel and fuses the rankings via Reciprocal Rank Fusion (k=60). The top-3 merged results are injected as <neuromcp-recall> context.

The hook is installed automatically by neuromcp-init-wiki and registered under UserPromptSubmit in Claude Code's settings.json. Re-run the indexer after large wiki updates (or schedule it — it's idempotent).

Tuning:

Known upstream issues

memories_vec does not reclaim space after DELETE — sqlite-vec #54 / #265. When you re-index after editing wiki sections, the old vector rows are marked deleted but their storage stays. The database file grows monotonically until you run npx neuromcp-index-wiki --rebuild, which drops and re-creates the vector rows. Run a rebuild every few weeks if you edit the wiki heavily.

claude CLI streaming hangs from non-TTY subprocesses on macOS — if you script interactions with claude -p from another process (e.g. scheduled jobs), pipe it through script -q /dev/null to allocate a pseudo-TTY. Without that the stdout buffer never flushes. We work around this inside the consolidator where needed.

Memory Governance

Namespaces isolate memories by project, agent, or domain.

Trust levels (high, medium, low, unverified) rank search results and control decay resistance.

Soft delete tombstones memories — recoverable for 30 days.

Content hashing (SHA-256) deduplicates at write time.

Lineage tracking records source, project ID, and agent ID per memory.

Configuration

All via environment variables. Defaults work for most setups.

What's New in v0.9

Auto-Capture (v0.9.0)

Session hooks automatically extract high-signal events — no manual store_memory calls needed:

Full Pipeline Auto-Capture (v0.9.1)

Auto-captured memories now go through the full store pipeline: dedup, contradiction detection, embeddings, entity extraction, and claims — via HTTP endpoint (POST /api/store). Falls back to raw SQL when HTTP is unavailable.

Contradiction resolution now has three tiers:

• Supersede (score > 0.5): old memory invalidated, new one takes over
• Coexist (score 0.35–0.5): both kept, linked via contradicts edge in knowledge graph
• Flag (score 0.3–0.35): reported for review

Explain Mode (v0.9.2)

Every search_memory result includes an explain field:

{
  "explain": {
    "source_trust": { "level": "high", "reason": "Directly provided by user" },
    "temporal_validity": { "currently_valid": true, "superseded_by": null },
    "contradictions": [{ "memory_id": "abc", "content_preview": "...", "resolution": "coexist" }],
    "claims": [{ "subject": "neuromcp", "predicate": "version", "object": "0.9.2" }],
    "confidence": { "retrieval_score": 0.016, "source_trust_score": 1.0, "overall": 0.85 }
  }
}

We publish all of this — schema versions, consolidation math, critic output, benchmark numbers with CIs — so you can audit exactly what the system remembers and how. If another local-first system publishes the same or better, link welcome.

Comparison

License

MIT