Memory Server API
Port: 48912 (internal)
The memory server runs as a separate process and handles all persistent memory operations. It is not intended for direct external access — the main server proxies memory-related requests.
Internal endpoints
The memory server provides endpoints for:
- Storing new conversation turns with timestamps and embeddings
- Querying recent context for LLM prompt construction
- Searching semantically similar past conversations
- Compressing old conversations into summaries
- Managing memory review settings
Storage backend
| Table | Purpose |
|---|---|
time_indexed_original | Full conversation history |
time_indexed_compressed | Summarized conversation history |
| Embedding store | Vector embeddings for semantic search |
Models used
| Task | Source |
|---|---|
| Embeddings | Local ONNX model bundled under data/embedding_models/<profile>/ (see memory/embeddings.py::EmbeddingService) |
| Fact extraction / signal detection / reflection / promotion merge / fact dedup / recall rerank | Tier summary (get_model_api_config('summary')) |
| History review / persona correction | Tier correction (get_model_api_config('correction')) |
| Negative-target keyword check | Tier emotion |
Communication
The main server communicates with the memory server via HTTP requests and a persistent sync connector thread (cross_server.py).