MemFit: Efficient Long-Term Agentic Memory

简体中文

L1 append-only and hybrid-retrieval hook diagnostic; no embedding model or cross-encoder is called.

Paper information

Field	Details
Paper	arXiv v1
Institution	The paper's first page does not list an affiliation for the first author.
First public release	2026-09-30 (arXiv v1)
Author code	No public author implementation found as of 2026-10-03.
Local adapter or method key	memfit
Local reproduction code	src/auto_research/agent_research/latest_20261003.py

Paper method and evidence

Append original turns without rewriting and use segment summaries only to build the index. Reads combine lexical, semantic, and reranking signals, avoiding expensive LLM writes.

flowchart LR
  T[Raw turns] --> W[Append-only store]
  W --> I[Index using segment summaries]
  I --> R[Hybrid retrieval and reranking]

Original paper figure

Figure 2 shows the memory pipeline. From the original paper; copyright remains with the authors.

Core equation

\(score=lexical+semantic\); writes are append-only.

Reported results

The authors report lower write latency and cost than LLM-based memory pipelines while retaining long-term QA quality.

Local reproduction

Compare with the mechanism disabled/default state. The three-seed diagnostic is diagnostic_only=true; no formal gain or capability ranking.