EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning

EverMemOS argues that memory systems fail less from missing facts than from failing to combine them. Its example: a user said last month they love IPAs and last week that a dentist prescribed antibiotics for two weeks. Asked for a drink for movie night, a snippet-based memory recalls the IPA preference and recommends beer. A memory that has consolidated the user's current state recommends a mocktail.

The system has three phases, loosely modeled on how biological memory is formed and recalled.

Episodic trace formation. An LLM watches the dialogue through a sliding window and cuts it at topic shifts. Each segment is rewritten as a concise third-person episode, from which the system extracts atomic facts (for precise matching) and foresight: forward-looking notes such as plans or temporary states, each with a validity interval ("taking antibiotics: two weeks"). Episode, facts, foresight and metadata form a MemCell.

Semantic consolidation. Each new MemCell is embedded and compared with existing MemScene centroids; above a threshold it joins the scene, otherwise it starts a new one. Scene summaries feed a compact user profile of explicit facts and implicit traits.

Reconstructive recollection. At question time, MemCells are scored by fusing dense and BM25 retrieval over their atomic facts; each scene takes its best cell's score; the top 10 scenes contribute episodes, which are reranked down to 10. Only foresight valid at the current time is kept. An LLM checks whether the context is sufficient and rewrites the query if not.

On LoCoMo and LongMemEval, EverMemOS outscored Zep, Mem0, MemOS and others.

Assistants deployed for a customer accumulate months of conversation. The useful pattern in this paper is turning that stream into three kinds of memory, each with its own role.

Episodes for context. A short narrative of what happened, rewritten so pronouns and references are resolved, is easier to retrieve and to read than raw turns.

Atomic facts for matching. Short verifiable statements make keyword and vector search precise; retrieval runs over facts, then returns the richer episode.

Foresight with expiry. Temporary states ("traveling next week", "on medication for two weeks", "budget frozen until Q3") are exactly what assistants get wrong later. Storing them with validity intervals and filtering by the current date stops stale constraints from steering answers, and stops expired ones from being forgotten too early, if the dates are extracted correctly.

Two practical warnings. First, structure is expensive to build. The paper's own log shows 7,056 LLM calls and 9.42 million tokens to build memory for 10 LoCoMo conversations of about 9,000 tokens each, around 100 tokens of processing per token of conversation. On a benchmark where each conversation fits in the context window, building plus retrieval costs more per question than simply including the conversation. The trade changes when histories are long, as in LongMemEval's 115,000 tokens.

Second, check the baselines. The "no memory" ablation scores 0.52% on LoCoMo, which suggests it received no history at all rather than the whole conversation; and the LongMemEval baselines come from a leaderboard, where Zep's row matches its run with a different, older answer model.