Memact Memory

Memact means act-on-memory.

Memory stores the accepted user memory that survives after Wiki review. Apps can suggest memory, Access checks permission, Context shapes it, Wiki lets the user accept/edit/reject/delete it, and Memory stores what remains useful.

Storage is local-first today, with a path for user-owned personal cloud storage later.

Owns

• Durable memory records.
• User memory records.
• Accepted Wiki entries.
• App-safe summaries.
• Corrections and forgetting records.
• Retrieval for apps and user views.
• CRUD operations for memory records.
• RAG-style retrieval context for allowed app reads.
• Field-based retrieval for approved memory fragments.
• Task context packets for memory-blind local workers.

Does Not Own

• Capture.
• Memory suggestion formation.
• Wiki moderation UI.
• API key verification.
• Full-Wiki access for apps.

Current Code

The v0 engine supports:

• createMemory(memoryInput, memoryStore)
• readMemory(memoryId, memoryStore)
• updateMemory(memoryId, patch, memoryStore)
• deleteMemory(memoryId, memoryStore)
• retrieveMemories(query, memoryStore, options)
• buildRagContext(query, memoryStore, options)
• buildTaskContextPacket(options)
• TemplateContextWorker.run(packet)
• rememberSchemaPacket(packet)
• rememberFeatureOutput(output) for compatibility with older feature output records
• retrieveContext(query, memoryStore, options)
• retrieveSchemaPackets(filter)
• buildContextForFeature(featureId, options)
• createCorrection(memoryId, correction)
• forgetMemory(memoryId, reason)

Summary retrieval returns compact records by default. RAG context is built from allowed memories, relation trails, and supporting snippets. Raw graph-style retrieval is a separate permission boundary and should not be treated as the default app response.

Task Context Packets

Memact workers must stay memory-blind by default.

For tasks such as onboarding prefill, field mapping, or context conversion, Memory can build a memact.task_context_packet.v0 packet. The packet includes only approved field fragments for one app and one connection. It filters out pending memory, unrelated categories, raw activity dumps, full profile fields, and sensitive fields that were not explicitly allowed.

The current worker is local and deterministic. It does not call OpenAI, Anthropic, Gemini, or any cloud model. Future providers can plug into the same worker interface, but they should still receive only the packet, not the full memory store.

See docs/task-context-workers.md.

Development

npm install
npm run check