Memory System

xbot supports pluggable memory providers. The agent uses memory to persist information across conversations, recall past interactions, and maintain user-specific context. The built-in default is xbot; flat and letta are alternatives, and none disables memory completely (no tools, no prompt).

# Set via environment variable / config.json
MEMORY_PROVIDER=xbot    # xbot (default) | flat | letta | none

none is an explicit mode: the agent receives no memory tools and no memory prompt at all — nothing is recalled, consolidated, or injected.

For Letta mode, configure an embedding model:

LLM_EMBEDDING_PROVIDER=openai
LLM_EMBEDDING_BASE_URL=https://api.openai.com/v1
LLM_EMBEDDING_API_KEY=sk-xxx
LLM_EMBEDDING_MODEL=text-embedding-3-small

All long-term memories are stored as a single text blob and injected into the system prompt on every request. When you start a new conversation (/new), the LLM consolidates and summarizes the memory blob.

Tools: memory_write, memory_list

Tip

Flat is dead simple — a good fit when you want a human-readable text blob and no SQLite. The built-in xbot provider is the default.

Three-tier memory architecture inspired by MemGPT. Powered by SQLite with vector embeddings and full-text search.

Three blocks always present in the system prompt:

Tools: core_memory_append, core_memory_replace, rethink

Long-term vector-backed storage for detailed facts, events, and context. Stored in SQLite with chromem-go embeddings for semantic search.

Tools: archival_memory_insert, archival_memory_search

Full conversation history searchable by date range. Powered by SQLite FTS5 (full-text search) for fast keyword matching.

Tools: recall_memory_search

Set memory_provider: "none" to run the agent with no memory capability at all:

• No memory_*, core_memory_*, archival_memory_*, or recall_memory_search tools are registered — the model never sees them.
• No memory section is injected into the system prompt.
• No recall injection, no automatic memory consolidation, and no compression memory hooks. SubAgents also get no memory.

Note

none is a deliberate mode, not a fallback. An unknown/unregistered provider name still gets the generic flat prompt guide; only none disables memory.

When starting a new conversation (/new command):

In multi-tenant deployments (Server mode with multiple channels like Feishu + Web):

• persona is global — the agent’s core identity is shared across all users
• human is per-user — what the agent learns about Alice is invisible to Bob
• working_context is per-session — each conversation maintains its own task context

Important

Core memory blocks set via core_memory_append/core_memory_replace are persistent across sessions. Use rethink to trigger the agent to re-examine and evolve its memory, similar to MemGPT’s self-editing mechanism.

• Built-in Tools — memory tools
• MCP — external tool integration
• Configuration — memory_provider setting