Overview

LLM conversations have a finite context window. Facts, preferences, and decisions made early in a conversation get pushed out when the window fills up, gets compacted, or a sliding window moves.

This package gives LLM agents persistent memory. It stores text chunks (facts, preferences, decisions, notes) and automatically injects the relevant ones back into conversations using semantic similarity.

The main capabilities:

• LLMs can save and recall memories via chat tools (add_memory, recall_memories, etc.)
• Memories auto-inject as synthetic tool calls when semantically related to conversation context
• Each memory has content, tags, summary, importance score, and a cached embedding vector
• Everything persists as JSON files on disk

Navigation

• Quick start: install and first use
• How it works: semantic recall, PageRank, auto-injection explained
• API reference: configuration, pool, memory fields
• Tool reference: tool parameters in detail
• Hook-based tool injection: synthetic tool calls from hooks
• Environment variables: configuration reference

License

MIT