Tencent-Hunyuan/Hy-Memory

The problem

Agents forget. Conversations reset, coding assistants lose the reasoning behind a patch, and every new session re-derives facts the user already stated. Bolting a vector store onto an app gives you raw similarity search, not memory: no conflict resolution, no distinction between a passing remark and a durable preference, and nothing tying a code revision back to why it happened. Hy Memory is a Python SDK plus memory pipelines that sit between your agent and its storage, writing structured memories and serving them back on query.

What's different

Most memory layers are a single retrieval index over past turns. Hy Memory splits by scene instead: an 8-layer hierarchy for chat (raw dialogue up to schemas and prospective intentions), a two-layer fact/experience split for working sessions, and a version-evolution diff-chain for coding that keeps code revision trajectories, not just text. It also ships a purpose-trained 30B-A3B MoE model for memory extraction and conflict reconciliation, released with FP8 and GGUF variants and a technical report. The dual-process framing (fast recall vs. deliberate reasoning) is the organizing idea, backed by an EMNLP 2026 paper. Compared with generic agent frameworks that leave memory to the developer, or standalone vector databases, this is opinionated about what a memory is.

Who it's for

Teams building assistant or coding-agent products that need memory to survive across sessions and deserve structured recall, especially those already storing conversations in Chroma, Qdrant, FAISS or Neo4j. It fits Python backends with an OpenAI-compatible LLM and embedding endpoint. Look elsewhere if you want a drop-in one-file memory class with no LLM calls in the write path, if your app is not Python, or if you only need short-context conversation buffering.

Cost of adoption

Requires an LLM endpoint and an embedding endpoint; the core install pulls ChromaDB and Kuzu, with optional extras for Qdrant, FAISS, Neo4j, MySQL and a local reranker. Memory writes can consume LLM calls, and the scene routing defaults differ between lite and pro/ultra modes, so behavior depends on configuration. The license badge says Apache-2.0 but the repository license field is NOASSERTION, which is worth checking before commercial use. The whole monorepo is days old; expect API churn.

Verdict

This is a brand-new project, nine days old, averaging roughly 133 stars a day and reaching #20 on the new Python board, with no Hacker News discussion yet. The research backing is unusually heavy for something this young, and the papers plus a released model suggest an organization investing for the long term rather than riding a spike. That said, one week of history is not stability: early adopters should pin versions and expect breaking changes. It is worth a serious evaluation if agent memory is core to your product; it is premature as an unexamined dependency.