Tencent Hunyuan open-sources Hy-Memory 2.0, a three-layer persistent memory stack for AI agents

Tencent's Hunyuan team has released Hy-Memory 2.0, a persistent-memory stack designed to give coding agents and conversational assistants durable knowledge retention, and has open-sourced the entire ecosystem under the Apache 2.0 license on GitHub and Hugging Face. According to a dev.to post detailing the announcement, the release bundles three parts: the hy-memory client library and SDK (stable version 2.0.2), a local management application called MemHub, and Hy-Memory-30B-A3B, a sparse mixture-of-experts model dedicated to memory extraction.

The problem it targets

The dev.to write-up frames the release as an answer to a structural weakness of large context windows: accumulating raw conversation history degrades attention accuracy, slows every dialogue cycle and drives up inference costs. Agents working through multi-week software projects, the post notes, routinely forget decisions made ten dialogue turns earlier or fall back into debugging loops over code that was already fixed. Off-the-shelf retrieval-augmented generation does not solve this, the author argues, because standard RAG has no sense of time: vectorially similar but outdated documentation, or contradictory early instructions, are retrieved with the same relevance score as a directive issued minutes ago.

A three-layer memory architecture

Hy-Memory 2.0 splits knowledge into three strictly separated layers — Working Memory, Code Memory and Chat Memory — each governed by its own retention rules, representation granularity and expiry policies. Working Memory acts as the agent's volatile scratch space, tracking the current plan, pending sub-tasks, raw outputs from recently invoked tools and intermediate reasoning. Rather than replaying the full history of a fifty-iteration refactoring loop, it synthesizes only the delta of current state: which files passed their unit tests, which functions failed to compile, and what step comes next. The remaining layers separate repository- and AST-level knowledge in Code Memory from conversational history in Chat Memory, according to the post.

MemHub and the memory model

MemHub ships as a desktop and command-line application with plug-and-play integrations for ten leading development environments, editors and autonomous assistants, aiming to act as a shared memory substrate across tools instead of tying retention to one editor. The extraction model, Hy-Memory-30B-A3B, is built on a sparse expert topology with 30 billion parameters in total and only about 3 billion activated per forward pass. It parses interaction streams to extract entities, detect changes in temporal premises and deduplicate information before consolidating it into a local SQLite relational store and hierarchical vector indexes. The published package includes the full weights, FP8 and INT4 quantizations for local workstation execution, and Python and TypeScript libraries for extending corporate pipelines.

Reported benchmark results

On LoCoMo, an academic benchmark for long-horizon conversational memory, the dev.to post reports a new global best of 72.6% accuracy, alongside a drop in the average number of words processed per query from 1,159 to 703 — a 39.3% cut in context overhead. Working Memory was evaluated separately on MemoryArena, a suite that measures whether agents stay consistent across 128 sequential cycles without hallucinating completed tasks; it scored 55.31%, which the post says is 8.4 percentage points ahead of pipelines that rely purely on sequential summarization.

Local-first positioning

The release leans heavily on privacy and decoupling. All context data, repository edit histories and dialogue logs stay on the developer's machine or on company infrastructure in encrypted SQLite structures and compact HNSW indexes, with no internal code traveling to third-party providers for summarization or consolidation. The memory engine is also decoupled from the reasoning model: developers can keep using their preferred foundation model while the auxiliary expert model quietly handles extraction, indexing and contextual injection, with the post claiming response times under five milliseconds for that cycle.

Why it matters

Long-horizon agents tend to fail on memory before they fail on reasoning, and this release treats memory as infrastructure rather than a prompt-engineering afterthought. An Apache-2.0, local-first stack with a purpose-built extraction model gives engineering teams an alternative to proprietary memory services and fragile homegrown RAG, without surrendering code or conversation data to a vendor. The headline numbers, if they hold under independent evaluation, point to better recall with roughly 39% less context per query — a direct saving in latency and inference cost. The larger architectural bet is that memory becomes a swappable, vendor-neutral layer beneath any reasoning model, which is precisely the piece most agent stacks currently rebuild badly in-house.