Persistent Memory

Persistent Memory

LingXiao has a two-layer memory architecture: long-term memory (FTS5 + BM25 full-text search engine) and short-term skill injection (execution knowledge injected at Worker dispatch time). Together they form a complete memory system.

Long-Term Memory

Full-Text Search Engine

Based on SQLite FTS5 + BM25 ranking:

• FTS5: SQLite full-text search extension with Chinese tokenization support
• BM25: Probabilistic relevance ranking algorithm, sorts by keyword match score
• Search interface: memory(action="search", query="keywords")

4 Memory Types

Type	Identifier	Description
User preferences	user	User’s coding style, tech preferences, work habits
Feedback guidelines	feedback	Agent working principles, best practices, lessons learned
Project knowledge	project	Project architecture, tech stack, key decisions, module relationships
External references	reference	Documentation links, API references, tool usage

Storage Locations

Scope	Path	Description
Project-level	.lingxiao/memory/	Project-specific memory, travels with the project
User-level	~/.lingxiao/memory/	Global user memory, shared across projects

Memory Operations

// Save memory
memory(action="save", name="user-preferences", type="user",
  description="User preferences", content="Prefers React + TypeScript...")

// Search memory
memory(action="search", query="frontend framework preferences")

// Load specific memory
memory(action="load", name="user-preferences")

// List all memories
memory(action="list", scope="project")

// Delete memory
memory(action="delete", name="outdated-memory")

Auto Distillation (Distill)

DistillCommand automatically distills valuable memories from session history:

• Scans completed sessions
• Identifies reusable knowledge, patterns, and lessons
• Auto-categorizes into 4 memory types
• Deduplicates before saving to the long-term memory store

Associative Memory (Dream)

DreamCommand performs associative memory maintenance during idle time:

• Discovers implicit associations between memories
• Merges duplicate or similar memory entries
• Updates status markers for outdated memories
• Optimizes search index performance

Maintenance Pipeline

The memory system has an automatic maintenance pipeline:

1. Index rebuild: memory(action="rebuild") rebuilds FTS5 index
2. Expiry cleanup: Automatically marks and cleans expired memories
3. Relationship updates: Cross-references between memories are auto-maintained
4. Capacity control: Project-level and user-level memory capacity monitoring

Short-Term Skill Injection

At Worker dispatch time, the Leader injects relevant Skill content into the Worker’s prompt based on task type and role bindings:

• Skill source priority: Project-level > Plugin-contributed > User-level > Built-in
• Only Skill phases relevant to the current task are injected
• Skill content serves as execution knowledge guiding the Worker’s workflow

This complements long-term memory: long-term memory provides cross-session knowledge accumulation, while short-term skill injection provides execution guidance for the current task.