Universal Memory-Import Tool for AgentBrain

"Bring your memory in 60 seconds. From anywhere."

Migrate existing memory from ChatGPT, Claude.ai, Gemini, Perplexity, or any LLM platform INTO AgentBrain. No lock-in. Your data, portable.

Why this tool exists

Cold-Start Problem (today, without this)

Customer signs up → workspace_id created → 0 memories. Cascading impact:

• substrate_quality_score: 0 (gate-blocks L3/L4 features per Drift-#21 substrate-gate)
• recall_pipeline: empty results
• bridge_summary cron: processed=0 every cycle
• dream_cycle first night: no patterns, no predictions
• llm_insight_generator: abstain (post Lane-I substrate-gate 2026-06-02)
• calibration_history: n=0, no signal
• identity_backbone bootstrap-v2: generic-seed only
• Time-to-aha-moment: weeks-to-months for organic substrate accumulation
• Result: free-tier signup abandons before reaching value

With this tool (Day-1)

Plus: defensive moat. "Your memory is portable, no lock-in with Anthropic/OpenAI." Hero-Claim Launch material.

Supported Platforms (V1)

V2 candidates (post-Launch): Notion AI, Mem.ai, Obsidian-with-AI-plugin, Cursor, Windsurf.

Architecture (high-level)

┌──────────────────┐
│  Customer upload │ (file: .zip / .json)
│  via CLI or UI   │
└────────┬─────────┘
         │
         v
┌──────────────────┐
│ Format-Detector  │ → sniff signature, select adapter
└────────┬─────────┘
         │
         v
┌──────────────────┐
│ Platform Adapter │ → parse → list[RawMemoryUnit]
│ (per platform)   │
└────────┬─────────┘
         │
         v
┌──────────────────┐
│ Quality-Filter   │ → LLM-classifier OR rule-based: memory-worthy ?
│ (optional)       │
└────────┬─────────┘
         │
         v
┌──────────────────┐
│ Normalizer       │ → list[CanonicalMemory] (Pydantic v2)
└────────┬─────────┘
         │
         v
┌──────────────────┐
│ Bulk-Importer    │ → batched POST /memory/store with provenance.imported_from=<platform>
└────────┬─────────┘
         │
         v
┌──────────────────┐
│ Migration-Report │ → log file + customer-facing summary (X imported, Y skipped)
└──────────────────┘

Surfaces

1. CLI Tool

# Auto-detect format
agentbrain-import --workspace <ws-id> --api-key <key> --file path/to/export.zip

# Force adapter
agentbrain-import --workspace <ws-id> --api-key <key> --file path/to/export.json --adapter anthropic-claude

# With quality-filter
agentbrain-import --workspace <ws-id> --api-key <key> --file path/to/export.zip --quality-filter haiku

# Dry-run (preview only, no writes)
agentbrain-import --workspace <ws-id> --api-key <key> --file path/to/export.zip --dry-run

# From Anthropic Memory Stores API (live pull, no file)
agentbrain-import --workspace <ws-id> --api-key <key> --anthropic-api-key <ant-key>

2. Backend Endpoint (optional v2)

POST https://api.agentbrain.ch/memory/import
Content-Type: multipart/form-data
X-API-Key: <agentbrain-key>

file: <upload>
adapter: anthropic-claude (optional, default: auto-detect)
quality_filter: haiku (optional, default: off)
dry_run: false

Returns: streaming import-progress + final report.

3. Customer-Facing Web-UI (Sprint-B Track-2 handles, not here)

Post-signup landing page CTA: "Import your existing memory" → upload-form → trigger backend endpoint.

Eigen-Use (before customer-launch)

Idee: vor customer-launch importiere mir Max + Thesh + Aura's eigeni Anthropic Claude.ai history INTO their canary workspaces. Result:

• Demos showe rich substrate (nöd 18-test-residue)
• Self-model has real expertise-domains
• recall_pipeline returns meaningful neighbors
• dream_cycle produces interesting patterns

Constraint: Privacy — only Max/Thesh/Aura workspaces, only Thesh-owned data, audit-trail per import.

Status

• SPEC: Ready (see SPEC.md)
• PLAN: Ready (see PLAN.md)
• Implementation: Pending coder-agent dispatch
• Target: ready by Sprint-A end (2026-06-12) OR earlier per re-priority discussion

Owner

Master Max (CTO). Implementation: TBD coder-agent. Coordination via inbox-from-master.md / outbox-to-master.md.