Gigabrain

Why this exists

Each AI assistant keeps a separate memory of your work, so one assistant can miss a decision that you made with another. You then explain the same facts again. Old information can also guide new work.

When a Codex session moves an app from Stripe to Paddle, Claude Code can still hold the old Stripe record. A later billing change can then use Stripe code.

Related MCP server: THOUGHT

What Gigabrain does

Gigabrain keeps one local record for supported AI assistants.

• Read supported memory files. Gigabrain adds the facts that it finds to the shared record. It keeps each source file unchanged.

• Keep the source and dates. You can check where a fact came from and when it entered the store.

• Show conflicts. Gigabrain applies documented rules when two records disagree about the same fact. It gives more weight to strong support and recent evidence, then keeps the decision history.

• Store data on your computer. The default memory store stays on your computer. You choose each optional connection.

Gigabrain works beside each assistant's built-in memory as a shared record that you can inspect or correct. Use an export when you want to move it.

Quickstart

These commands install the package, connect a project, check the setup, and write an audit report. All supported imports stay read-only.

npm install @legendaryvibecoder/gigabrain
npx gigabrainctl init --project-root /path/to/repo
npx gigabrainctl doctor --config ~/.gigabrain/config.json --target both
npx gigabrainctl handoff --config ~/.gigabrain/config.json \
  --output-dir ./gigabrain-memory-audit

init writes the configuration to ~/.gigabrain/config.json. The other commands and assistant setups use this file. Read the audit report before you enable a feature from the configuration guide.

Open the guide for your assistant: Codex · Claude Code · OpenClaw.

Limits and safeguards

• Memories in ChatGPT, Claude.ai, Gemini, and Copilot stay in those accounts. Gigabrain can import a supported file after you export it.

• A laptop and a desktop use separate stores. Use a reviewed bundle or a configured connection when you want to move data between them. See the sharing guide.

• Gigabrain ranks evidence with documented rules and lets you inspect or change the result.

• The local store can contain sensitive data. Protect the computer and review every export before you move it. Read the privacy guide.

The configuration guide describes every setting and optional feature.

Technical reference

Core behavior

• Shared recall with clear scope: Codex, Claude Code, Hermes, and OpenClaw can share project facts and stable user preferences when they use the same Gigabrain configuration.

• Source record: Each claim keeps its source host, path, evidence, time data, status, and trust tier.

• Conflict rules: Gigabrain checks source trust first. It then checks independent support. Recency resolves the final tie. An append-only record keeps each decision.

• Time model: A fact can have a content time and a validity window. Normal recall skips expired or superseded rows.

• Audit tools: Optional tools can recover transcripts, filter secret risks, manage review queues, write Handoff Records, and build a memory wiki.

• Local operation: The default setup runs locally, where SQLite stores the data and lexical search finds it. Policy checks control each result. A cloud memory service is optional.

Processing flow

1. Ingest: Add supported local memory files, explicit checkpoints, and manual exports to a local event store.

2. Build the current view: Write the latest state to SQLite and keep its source and validity data.

3. Resolve conflicts: Apply reviewable rules to duplicate or conflicting claims. A queue holds uncertain changes that can remove data.

4. Serve recall: Use lexical search through MCP, the CLI, or authenticated HTTP. Loopback Ollama embeddings are optional.

   Codex   Claude   Cursor   OpenClaw   Hermes   manual exports
     └────────┴────────┴─────────┴─────────┴──────────────┘
                              │   (supported imports are read-only)
                              ▼
                    ┌────────────────────┐
                    │      Gigabrain     │   capture · resolve · recall
                    │ arbitration ledger │   MCP · CLI · HTTP
                    └─────────┬──────────┘
                              │  recall with sources after conflict checks
                              ▼
                  any configured agent uses gigabrain_recall

Release highlights

The cloud inbox, transcript recovery, Git wiki, Obsidian reference set, remote bridge, and URL importer start disabled. Read the configuration guide and privacy model before you add a data source or network provider.

Supported clients

Privacy model

• Gigabrain keeps the standalone SQLite store under ~/.gigabrain/. A hosted Gigabrain service is optional.

• The default LLM provider setting is none. Optional semantic embeddings go only to a loopback Ollama endpoint.

• Gigabrain extracts raw transcripts through a local provider or a local hook that you add. A cloud audit excludes rows that can contain credentials. Before data leaves the computer, Gigabrain masks supported PII patterns and removes the original scope.

• Native host stores use read-only import. To add a supported memory file from a cloud account, export it first.

• Enable each network feature before use. The remote bridge is opt-in. For URL import, the Python console requires an exact host allowlist and an explicit setting.

• The release gate scans all publishable files, the npm package contents, Git data, and GitHub data. Its output hides matched values.

The SQLite store and generated Markdown can contain sensitive memory. Protect the host account, use disk encryption, restrict file permissions, and review each export before you move it. Read the full boundary in the privacy model.

How it works under the hood

Conversation (OpenClaw / Codex / Claude Code / Claude Desktop)
               │
               ▼
┌──────────────────────────────────┐
│           Gigabrain              │
│   (memory layer + MCP server)   │
├──────────────────────────────────┤
│  Capture ─► Policy ─► Registry  │
│  Recall  ◄─ Orchestrator        │
│  Host Sync ◄─ Codex/Claude/etc. │  ← read-only adapters
│  Transcript Harvester ◄─ logs   │  ← optional recovery
│  Arbiter (trust>support>recency)│  ← at ingest + nightly
│  World Model (entities/beliefs) │
│  Git Wiki ◄─► reviewed edits    │  ← optional, high trust
└──────────────┬───────────────────┘
               │
         SQLite + FTS5 + optional local embeddings

• Capture: Explicit remember calls, checkpoints, host imports, and optional transcript recovery become append-only events.

• Recall: FTS5 and BM25 search text directly. When local embeddings are available, Gigabrain combines lexical and dense rankings. It then applies scope, status, answer fit, source data, and conflict rules.

• Conflict review: The world model records competing beliefs for each claim slot and ranks them by trust tier. It next checks independent support. Recency resolves any remaining tie. Clock-skew and source-independence checks protect the result.

• Audit reports: Static Markdown, HTML, and JSON reports show readiness, source coverage, conflicts, stale rows, and omitted secret risks.

The published evidence comes from a small development regression set. The benchmark evidence page describes its limits.

Why use it when native memory exists?

Several assistants provide useful native memory. Codex has local memories, Claude Code has auto memory, Cursor has project memories, and OpenClaw has hybrid memory search. Each product documents a different scope. Some stores stay on one machine or inside one product.

Gigabrain keeps source data across products and separates project facts from user preferences. Fixed rules resolve conflicts. Validity dates control when each fact applies. You can move the record with an export and audit the protocol independently. The detailed comparison maps these features to native memory.

MCP tools

gigabrain_recall · gigabrain_remember · gigabrain_checkpoint · gigabrain_provenance · gigabrain_recent · gigabrain_sources · gigabrain_sync_status · gigabrain_export_brief · gigabrain_entity · gigabrain_relationships · gigabrain_contradictions · gigabrain_arbitrate · gigabrain_adjudications · gigabrain_beliefs_as_of · gigabrain_review_queue · gigabrain_doctor

The coverage matrix lists each action and its available interface.

CLI

npx gigabrainctl init                       # Find and connect installed agents
npx gigabrainctl handoff --output-dir ./out # Write a memory audit and safe Handoff Records
npx gigabrainctl nightly                    # Run the nightly import, conflict check, and audit
npx gigabrainctl doctor                     # Check system health
npx gigabrainctl inventory                  # Show memory statistics
npx gigabrainctl review contradictions      # Review conflicts across agents
npx gigabrainctl sync-hosts --host codex,claude_code  # Import a host again
npx gigabrainctl vault sync|status          # Use the read-only Obsidian reference set
npx gigabrainctl transcript sync|status     # Recover raw local transcripts
npx gigabrainctl wiki project|reconcile|status  # Manage the Git versioned memory wiki
npx gigabrainctl watch --install-hook --kind=session  # Capture at the end of a session
npx gigabrainctl export-bundle --out ./memory-bundle.json
npx gigabrainctl import-bundle --in ./memory-bundle.json
npx gigabrainctl migrate legacy-drop --dry-run  # Preview legacy cleanup

All commands accept --config <path> and are also available as npm run scripts.

HTTP endpoints

/gb and /gb/health return service status only. Every data route accepts X-GB-Token, X-OpenClaw-Token, or a Bearer token. A route denies access when its configuration has no token. The dangerous development setting GB_ALLOW_NO_AUTH=1 can bypass Node route checks only when no token is configured. It prints a warning. Use this setting only in a disposable loopback environment. OpenClaw gateway authentication can add an outer layer.

Key subsystems

Prerequisites

• Node.js 22.18.0 or later. Gigabrain uses node:sqlite and built-in TypeScript type stripping.

• Ollama is optional. It provides local fact extraction and semantic search.

• OpenClaw 2026.2.15 or later is required for the plugin path.

• Python 3.10 or later is required for the optional web console.

Testing

node tests/run-all.js     # Run the repository tests
node scripts/package-smoke.js  # Test the packaged runtime
npm run pack:dry-run     # Check the published package contents
node scripts/check-no-pii.mjs
node scripts/check-public-mirror.mjs --require-single-commit  # Check the public mirror
npm run audit:github-metadata -- --repo owner/repository     # Check a new remote

Security

• Node HTTP endpoints that carry data require a token by default. The code compares tokens with timing-safe logic. A Node token grants access to its configured store, where scope filters recall and queries. The optional FastAPI console isolates scoped tokens and conceals whether an ID exists. Limit GB_ALLOW_NO_AUTH=1 to the development use described above.

• The optional web console listens on loopback as documented. It sets security headers and limits uploads. Extracted PDF text also has a size limit. URL import starts disabled.

• Release checks run npm audit and pip-audit against dependencies. The security review gives the date, result, and remaining risks.

• An explicit allowlist creates the public mirror with a new single-commit history. The private engineering repository stays private.

Report vulnerabilities through the private process in SECURITY.md.

Contributing

Read CONTRIBUTING.md before you send a change. Report a concrete bug in Issues, or use Discussions for design and usage questions. Remove private data from every post, including secrets, local paths, identifiers, and runtime files.

License

MIT License. See LICENSE.