Persistent agent memory that lives in plain text

What it does ReMe is a memory management framework for AI agents that compresses sprawling conversations to fit context windows, then persists the important bits across sessions. It tackles the usual agent amnesia—truncated history and stateless restarts—by automatically summarizing dialogue, caching tool outputs, and recalling relevant prior context through hybrid search. The file-based ReMeLight system keeps everything in a working directory of Markdown journals and JSONL logs you can read with cat.

The interesting bit Rather than burying memory in a vector database, ReMe treats your agent’s recollection like a notebook: a root MEMORY.md for persistent facts, daily journal files, and raw dialog archives you can edit or version-control. The framework delegates compaction to its own ReActAgents, and the sample log shows a 99.5% token reduction—223,838 tokens compressed to roughly 1,105—while retaining searchable semantics via vector-plus-BM25 retrieval.

Key highlights

• File-first architecture stores long-term memory as editable Markdown and structured JSONL instead of opaque database rows
• pre_reasoning_hook bundles context compaction, tool-result truncation, and summary generation before inference
• Hybrid memory_search pairs embeddings with BM25 for semantic and exact-match recall
• Async summarization writes daily memory files in the background without blocking the conversation loop
• Claims state-of-the-art results on the LoCoMo and HaluMem benchmarks

Verdict ReMe suits developers building agents that need inspectable, persistent memory without managing a separate database. If you prefer your context store to be a black box—or if you need a fully offline, API-free stack—look elsewhere.