# WhiteMagic — governed local memory for agent continuity

> WhiteMagic — governed local memory for agent continuity is an open-source MCP server for knowledge and memory. Local-first memory for coding agents with signed continuity receipts. It can run locally on your machine or as a hosted remote server.

- **Kind:** MCP server
- **Pricing:** Open source
- **Platforms:** CLI, Linux, macOS, Windows, self-hosted, API
- **Website:** https://www.whitemagic.dev
- **Repository:** https://github.com/lbailey94/whitemagic/tree/HEAD/npm/whitemagic-mcp
- **MCP Registry name:** io.github.lbailey94/whitemagic-mcp
- **Categories:** knowledge and memory, developer tools
- **Listing updated:** 2026-11-14

## About

WhiteMagic is a single Rust binary, `wm`, that the npm package downloads and checksum-verifies on first run, then launches as an MCP server. It stores project notes, decisions and session checkpoints on your machine so the next session, restart or different client can pick up where you stopped.

The site organises it in three arcs: memory (on-device holographic memory and optional client-encrypted Memory Crystals), containment (Linux Landlock filesystem and network confinement) and provenance (Ed25519-signed continuity receipts, checkable offline). The local core is MIT licensed, needs no account and sends no usage telemetry by default.

Separately, a hosted lane offers read-only recall over a public corpus with a free evaluation key at 50 recalls a day. The stable release shown is 11.2.1; an experimental channel with Landlock sandboxing keeps its own version and store path.

## Features

- **Local memory store:** Keeps notes and decisions on your hardware.
- **Session checkpoints:** Resume work across restarts and clients.
- **Signed receipts:** Ed25519 continuity receipts verifiable offline.
- **Landlock sandboxing:** Kernel-level filesystem and network confinement (experimental channel).
- **Memory Crystals:** Optional client-encrypted memory.
- **Hosted recall:** Free evaluation key for read-only recall, 50 a day.

## Good for

- Resume a coding project in a new session without re-explaining
- Share one memory between Claude Code and Cursor
- Hand off work to a peer agent with a verifiable record
- Try hosted recall without installing anything

## Install

### Claude Code (Terminal)

```sh
claude mcp add --transport http whitemagic-mcp https://mcp.whitemagic.dev/mcp
```

### Claude Desktop (Settings → Connectors → Add custom connector)

```sh
https://mcp.whitemagic.dev/mcp
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "whitemagic-mcp": {
      "url": "https://mcp.whitemagic.dev/mcp"
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "whitemagic-mcp": {
      "type": "http",
      "url": "https://mcp.whitemagic.dev/mcp"
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "whitemagic-mcp": {
      "serverUrl": "https://mcp.whitemagic.dev/mcp"
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "whitemagic-mcp": {
      "type": "http",
      "url": "https://mcp.whitemagic.dev/mcp"
    }
  }
}
```

## Tools

- `wm`: Hosted router for available hosted tools. Use tools/list to discover supported routes.
- `memory.create`: Store a new memory in the holographic substrate with bijective UUID, 6D projection, and galaxy routing. (READ-ONLY: mutations refused)
- `memory.search`: Search memories: hybrid BM25 + vector fusion with deterministic sub-millisecond retrieval.
- `memory.temporal_query`: Query scoped temporal source assertions and their statuses; results are evidence, not verified facts.
- `memory.read`: Retrieve a specific memory by its numeric record ID or UUID string.
- `memory.list`: List memories with pagination (limit, offset) and galaxy filtering.
- `memory.hybrid_recall`: Compatibility alias of memory.search with BM25 + dense vector fusion.
- `galaxy.list`: List all active sovereign memory galaxies, their record counts, sample tags, and starter guide status.
- `galaxy.fork`: Fork an existing galaxy partition into a new sovereign galaxy branch (e.g. 'guide' -> 'project-alpha'). (READ-ONLY: mutations refused)
- `galaxy.create`: Establish a new sovereign galaxy namespace with an initial genesis beacon. (READ-ONLY: mutations refused)
- `session.record`: Record a session turn (decision, breakthrough, summary) for compounding continuity. (READ-ONLY: mutations refused)
- `session.continuity`: Recall where the previous session left off: recent turns, salient anchors, unresolved DiAML arcs, related track crystals, checkpoints, and open queues.
- `session.start`: Start or resume an agent session for persistent continuity across tool invocations.
- `memory.update`: Update an existing memory by writing a superseding record (or consolidating into an existing record) and a durable revision entry (prior/new content hashes, timestamp, reason). (READ-ONLY: mutations refused)
- `memory.revisions`: List the durable revision chain of a memory, or verify it by recomputing content hashes and chain links.
- `session.checkpoint`: Save a structured checkpoint (summary, git state, next queue, open flags, track/lease metadata) for lossless handoffs. (READ-ONLY: mutations refused)
- `memory.ingest`: Ingest inline items (items / items_jsonl / text) or a JSONL/text file (source) via the batch-create path; returns real counts and refuses malformed input. (READ-ONLY: mutations refused)
- `receipts.emit`: Emit a signed continuity receipt: kind 'session' signs a digest of the last `limit` session-log turns; kind 'state_transition' signs a store-state snapshot chained to the prior emitted receipt. 'task' receipts are not supported.
- `receipts.verify`: Verify one supported WhiteMagic receipt profile (including emitted session/state_transition continuity receipts) against the existing local gate key. This is not a general Continuity Receipt bundle verifier.
- `memory.count`: Count memories matching optional filters without returning full payload content.
- `memory.stats`: Substrate statistics: epoch, total records, relations, postings, and crash-consistency invariants.
- `memory.tags`: List the tag vocabulary with usage counts.
- `memory.aggregate`: Aggregate over the matched record set: numeric fields (importance, created_at, epoch) support count/sum/avg/min/max; categorical fields (source, domain, class, status, galaxy) support count only. Unsupported field/op is refused.
- `memory.associations`: Traverse associative graph neighbours of a memory over hub-penalized multi-hop spreading activation.
- `memory.batch_read`: Read multiple memories by their UUID or numeric record IDs in a single call.
- `memory.query`: Structured query over memories with filters, ordering, and field projections.
- `memory.filter`: Filter memories by metadata (tags, minimum importance, recency).
- `memory.pin`: Durably pin a critical memory against decay, sweep compaction, and forgetting, or release it with pinned=false. (READ-ONLY: mutations refused)
- `memory.search_batch`: Execute multiple memory search queries concurrently with sub-millisecond retrieval.
- `session.list`: List session lanes with turn count, last timestamp, latest preview, start-marker flag, and checkpoint presence.
- `session.recall`: Recall specific interactions and turns within a targeted session (optional content query and limit).
- `session.replay`: Replay the chronological trajectory of a session: 'full' returns all turns, 'selective' filters by min_importance/turn_types, 'progressive' keeps the most recent turns within token_budget, and 'hybrid' combines earlier high-salience anchor turns (up to 30% of token_budget) with the most recent turns.
- `session.digest`: Compose a digest from typed session-log turns and the latest checkpoint (markdown text plus JSON, with optional codex crystallization).
- `mandala.status`: Status of the Mandala microsecond Landlock sandbox, ABI level, POSIX rlimit bounds, and capability ledger.
- `mandala.triage`: Non-autoregressive System 1 JEV/Layla triage gate (<30 ns) to classify tasks and predict marginal utility and confinement tier.
- `mandala.evaluate`: Execute a candidate mutation in an ephemeral Landlock sandboxed Kekkai and notarize with Spec 0.5 receipt. (READ-ONLY: mutations refused)
- `mesh_sync`: Synchronize the Substrate with a remote peer node over the sovereign TCP protocol.
- `gnosis.status`: System epistemics, self-model calibration, and invariant health status.
- `gnosis.explain`: Explain internal architectural invariants, memory schemas, or routing logic.
- `decision.shortlist`: Rank candidate routes/actions for a state with the local static-embedding retrieval organ (no network, no model forward pass) and emit a signed shortlist receipt. (READ-ONLY: mutations refused)
- `decision.deliberate`: Deliberate over ambiguous candidate options using local grammar-constrained SLM inference (100% schema guarantee, no hallucination) and emit a signed deliberation receipt. (READ-ONLY: mutations refused)
- `decision.outcome`: Sign and journal the outcome of a dispatched decision/shortlist receipt (success, failure, corrected) for the learning loop. (READ-ONLY: mutations refused)
- `sutra.cast`: Evaluate and dispatch a zero-allocation Taoist Hand Seal / Sutra Codon (e.g. '觀 證 行', 'Kek-01-Seal'). Validates preconditions, computes Kekkai sandbox boundaries, and executes bound Action Skeleton cycles or transmits swarm mudra packet to P2P peer. (READ-ONLY: mutations refused)
- `sutra.compile`: Compile high-level natural language intent into an optimal compound Sutra codon and Taoist Mudra formula using static Model2Vec embeddings or lexical centroids.
- `sutra.canon`: Inspect the canonical 12-glyph logoglyphic Sutra canon, alchemical dual-mappings, Taoist hand seals (印诀), and bound Action Skeletons.
- `sutra.list`: List all recognized canonical compound Taoist Mudras (印诀式), their formulaic compositions, preconditions, and primary execution skeletons.
- `sutra.hexagram`: Inspect or mutate one of the 64 King Wen I Ching Hexagram Macro-Actions. Returns King Wen number, binary symbol, trigrams, Wu Xing element, default codon, nuclear inner, complement (反), reverse (覆), and dynamic line mutation transformations.
- `sutra.route`: Route an inquiry or task intent through the Catuṣkoṭi K3 dialectical cognitive router. Employs Kleene 3-valued logic to resolve dialectical tensions (competing action vectors like execute vs dry-run) and matches 64-hexagram macro-actions, returning System 0/1/2 triage tier, resting potential, plasticity, and bound Action Skeleton.
- `sutra.contract`: Resolve or verify bounded caller-asserted temporal assertions. Results describe the supplied metadata; this tool does not authenticate sources or enforce native-store sanctuary access.
- `legion.summon`: [Experimental / Simulation] Run a deterministic Rayon stack-codon cohort simulation across CPU cores. Evaluates N candidate codon variants across four archetypes (Adversarial, Refactor, Speculative Replay, Horizon Stress) and classifies heuristic outcomes via Cho-Han (丁半). Scores and token-budget deltas are modeled estimates, not measured LLM execution.
- `chohan.settle`: [Experimental / Simulation] Classify a payload or proof string into Cho (丁 / even / pass) or Han (半 / odd / anomaly). Advisory string/parity classification only; does not independently verify execution proofs or perform on-chain escrow.
- `valkyrie.pulse`: Execute a 12-Phase Tricameral Valkyrie-2 pulse (`Valkyrie-0` 1.58-bit Ternary + `RWKV-7` Delta-Rule 28-Gana Router + Cross-Layer Sparse Transcoder Attribution Graph + `Valkyrie-1` Corpus Callosum Fusion + Tsao 15 Hz Working-Memory Ring + Citadel Antibody Gate + Hardware Thermal Telemetry), or run the full real-world corpus & world model benchmark (`benchmark=true`).
- `service.request_eval_key`: Request a free evaluation API key in-band. Keyd applies persistent per-client-IP and global daily issuance limits. The returned key's recall and optional write-sample terms are read from the authoritative key record. Pass it as 'Authorization: Bearer <key>' on subsequent calls.
- `economy_advance`: Request a small uncollateralized compute advance. Phase 1 (2026-10-15): bound to your authenticated evaluation key (daily per-IP and global caps); new key identities start at a $0.10 ceiling that grows $0.25 per verified settled payment. Returns a recall-scope session pass; hosted write requires a paid lease.
- `economy_settle`: Settle outstanding compute balance to upgrade Bayesian reputation score and dynamically expand credit ceiling. Supports payment reference or gasless EIP-3009 transfer authorization.
- `economy_ledger`: Inspect public transparency metrics: total loans disbursed, default rates, zero predatory liquidation policy, and sovereign free invariants.
- `economy_chohan_arbitrate`: Advisory contention tie-breaker only. It does not acquire or enforce locks, uses a post-hoc recomputable SHA-256 draw, and is not host-bias-proof.
- `economy_chohan_roll`: Advisory SHA-256 dice draw. The server seed is disclosed with the result; there is no prior commitment or host-bias proof.
- `citadel_threat_intel`: Query zero-cost, real-time threat intelligence and hostile-source indicators captured by the Citadel deception mesh. Pass 'lookup' (an IP address, subnet prefix, or SSH HASSH fingerprint) for target analysis, or omit 'lookup' for the latest rolling attack telemetry, campaign clusters, and advisory subnets.

## FAQ

### Does it send telemetry?

The site says the local core sends no off-device usage telemetry by default.

### What licence?

The local core is MIT.

### How do I run it?

`npx -y whitemagic-mcp@11.2.1 serve --profile cyberbrain`, per the README.

### Which clients work?

Claude Code, Claude Desktop, Cursor, Codex, VS Code, opencode, Antigravity and Devin Desktop are named.

### Is WhiteMagic — governed local memory for agent continuity free?

Yes. WhiteMagic — governed local memory for agent continuity is open source, so you can use and self-host it for free.

### How do I add WhiteMagic — governed local memory for agent continuity to Claude Code?

Run this command in your terminal: claude mcp add --transport http whitemagic-mcp https://mcp.whitemagic.dev/mcp

### What are the best alternatives to WhiteMagic — governed local memory for agent continuity?

The closest alternatives to WhiteMagic — governed local memory for agent continuity are REA, Codebase Memory, GitHub, Serena MCP: the IDE for your agent and XERJ.

## Alternatives

- [REA](https://ai-powered.com/mcp/rea): REA is an open-source MCP server for developer tools.
- [Codebase Memory](https://ai-powered.com/mcp/codebase-memory): Codebase Memory is an open-source MCP server for knowledge and memory.
- [GitHub](https://ai-powered.com/mcp/github-2): GitHub is an open-source MCP server for developer tools.
- [Serena MCP: the IDE for your agent](https://ai-powered.com/mcp/serena-mcp-the-ide-for-your-agent): Serena MCP: the IDE for your agent is an open-source MCP server for developer tools.
- [XERJ](https://ai-powered.com/mcp/xerj): XERJ is an open-source MCP server for knowledge and memory.
- [Ouroboros](https://ai-powered.com/mcp/ouroboros): Ouroboros is an open-source MCP server for developer tools.
- [Claude Code Ultimate Guide](https://ai-powered.com/mcp/claude-code-ultimate-guide): Claude Code Ultimate Guide is an open-source MCP server for ai models and llm tooling.
- [Nautilus Compass Memory](https://ai-powered.com/mcp/nautilus-compass-memory): Nautilus Compass Memory is an open-source MCP server for knowledge and memory.

---
Source: https://ai-powered.com/mcp/whitemagic-governed-local-memory-for-agent-continuity
