AI assistants adding persistent memory

In July 2026, coverage highlighted growing interest in AI systems that can retain information over time. An initial overview described a variety of memory‑enabled AI products, separating user‑oriented tools such as Notion AI and Lindy from developer‑focused infrastructure that lets agents preserve state across sessions. It also emphasized that well‑structured documentation—hierarchical tables of contents, clear directory naming, and comprehensive READMEs—helps AI agents locate needed specifications efficiently.

The following day, reporting focused on concrete implementations by major AI developers. Google’s Gemini Gems, Anthropic’s Claude Projects, and Microsoft’s Copilot Agents introduced features that allow large‑language‑model assistants to keep user‑specific instructions and linked files across interactions. A separate project, Memoria, showcased a three‑tier memory architecture (Redis for short‑term, PostgreSQL with vector embeddings for personal memory, and a full‑transcript archive) that automatically decays, consolidates, and resolves contradictory data, achieving a 77.6% boost in decision‑making accuracy on benchmark tests. Memoria was developed for the Qwen Cloud Hackathon and utilizes Alibaba Cloud infrastructure.

By August 2026, the field expanded with new specialized tools and evaluation standards. Engram emerged as a standalone Go binary using an SQLite backend to provide coding agents with persistent memory, allowing for updates via MCP servers and shared learnings through Git. To standardize the growing field, the Agent Memory Leaderboard (AML) was established to isolate memory system performance from the influence of specific models or prompts. As of mid-August, the AML includes participation from 136 teams and 67 memory frameworks across open-source and commercial tracks.