Seme
AI companions forget. Seme remembers what you told them, links it up, and fixes it when it changes.
InfrastructureRepository
The Problem
Most companions either stuff the whole chat history into the prompt or quietly drift: stale facts linger, contradictions pile up, and nobody can say where a memory came from. Seme keeps every memory traceable and retires what is no longer true.
The Solution
A local-first memory system for AI companions. It extracts facts, preferences, routines and boundaries from conversation, stores them with their source turn in a single SQLite file, compiles them into wiki pages, and retrieves with vector and graph search.
What it does
- Extraction into typed memories: facts, preferences, routines, interests, relationships and boundaries
- Entity and relationship graph built alongside the memories
- Wiki pages compiled from memories, with draft, stable and canonical maturity
- Hybrid vector and graph retrieval for each question
- Dashboard for browsing memories, entities, the graph, wiki and tasks
What took work
- Superseding contradicted memories without losing the history of what was said
- Keeping vector search, graph and full-text in one SQLite file with sqlite-vec
- Swapping embedding providers (Ollama, OpenRouter, none) without re-architecting storage