지식 및 메모리 관리 v0.0.2: 휴대용 $ AGENT_HOME으로 릴리스 정리

We are releasing v0.0.2 of Knowledge-and-Memory-Management, focused on production-ready knowledge collection and memory management for agent-based systems. The headline feature is the full replacement of all personal paths with the portable $AGENT_HOME variable, eliminating configuration drift across environments.

What Changed for Practitioners

Previous versions required absolute paths in configuration files or code, which broke when moving between local dev, CI, and production nodes. Now every internal path respects $AGENT_HOME, and the default value is set to ~/.agent_home if unset, but you can override it in a single location. This means you can share agent configurations verbatim across a team.

Knowledge Collection from Diverse Sources

The core collectors for web, video (transcript + metadata), and long-form articles are now stable. Each source is ingested into a consistent schema: text content, source URL, title, and timestamp. The collectors handle HTML stripping, YouTube transcript retrieval, and PDF/paywall article parsing respectively.

For example, to collect from a web page:

import os
from agent_knowledge import Collector

# AGENT_HOME is typically exported via .env or CI secrets
os.environ["AGENT_HOME"] = "/opt/agent-data"

collector = Collector(agent_home=os.getenv("AGENT_HOME"))
# Ingest a web article – automatically chunked and embedded
collector.from_web(url="https://example.com/api-docs", tags=["api", "v2"])

Enter fullscreen mode Exit fullscreen mode

This produces a local vector store under $AGENT_HOME/knowledge/web/ along with metadata. The same interface works for video (collector.from_video(youtube_url=...)) and articles (collector.from_article(local_path=...)).

Memory Management Layer

Memory is now structured with three tiers:

• Episodic (recent interactions, auto-expire)
• Semantic (collected knowledge with embeddings)
• Procedural (reusable prompts and workflows)

Each tier respects $AGENT_HOME. In v0.0.2 , the memory manager supports semantic search using cosine similarity across all collected sources. You can query for context retrieval without exposing storage backends.

Portability in Practice

All file I/O (knowledge stores, index caches, configs) now uses Path(agent_home) / ... instead of hardcoded paths. This is critical for Docker containers where the home may be mounted at runtime, or for multi-instance agents that share a network drive.

What’s Next

The release is intentionally minimal—no over-engineered RAG pipelines yet. v0.1.0 will add time-aware memory decay and cross-source deduplication. For now, this gives you a solid foundation to experiment with knowledge ingestion and retrieval in a portable, production-happy format.

Check the repo for full docs and migration notes.