brain

Hy-Memory

Properties11
statusseedling
typesource
created2026-10-17
updated2026-10-17
tagsoss
aliases
sourcehttps://github.com/Tencent-Hunyuan/Hy-Memory
docs_urlhttps://github.com/Tencent-Hunyuan/Hy-Memory/wiki
languagePython
stars316
public

8 min read

Tencent Hunyuan

Hy Memory

Production-grade dual-system cognitive memory for LLM agents.

Latest News Papers Get Started System Architecture Support Hugging Face Model MemHub Website Discord Apache-2.0 License

English | δΈ­ζ–‡

Hy Memory (Official Website) is a production-grade dual-system cognitive memory engine and Python SDK (hy-memory) for self-evolving LLM agents, providing specialized memory pipelines across three task scenes: hierarchical long-term cognitive memory for Chat, two-layer fact and reusable-experience memory for Working, and version-evolution diff-chain tracking for Coding. Built on top of Hy Memory, MemHub is a local-first desktop and CLI product that unifies conversation history and coding traces from user-authorized AI applications (such as Claude Code, Cursor, Codex, OpenCode, and Qwen Code) into a single managed runtime and local data directory, enabling governed cross-application memory recall.

memhub_h265_10mb.mp4

MemHub Demo Video

Papers

Welcome to the Hy Memory Research Family, covering cognitive chat memory, long-horizon coding memory, and RL-trained memory writing models:

Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents (Accepted at EMNLP 2026)
TL;DR: Proposes the dual-process (System 1 + System 2) cognitive memory architecture for Chat, organizing personal memory into an 8-layer hierarchy (raw dialogues, atomic facts, summaries, schemas, and prospective intentions) with conflict reconciliation and cross-domain schema induction.
πŸ“„ Paper: https://arxiv.org/abs/2606.09483

CMB: A Coding Memory Benchmark for Long-Horizon and Interleaved Coding Agents (Submitted to ICLR 2027)
TL;DR: Introduces a benchmark and two-layer + version-evolution diff-chain memory method for long-horizon Coding agents, preserving both reusable engineering experiences and exact code revision trajectories across interleaved tasks.
πŸ“„ Paper: docs/papers/CMB_A_Coding_Memory_Benchmark_for_Long_Horizon_and_Interleaved_Coding_Agents.pdf

Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing (Open Model Technical Report)
TL;DR: Trains a 30B-A3B MoE model via attributed reinforcement learning specifically for high-precision long-term memory extraction and conflict reconciliation.
πŸ“„ Paper: https://arxiv.org/abs/2610.22196 Β· Model: https://huggingface.co/tencent/Hy-Memory-30B-A3B

Quick Start

1. Installation

pip install hy-memory==2.0.2              # core (includes ChromaDB and Kuzu)
pip install "hy-memory[qdrant]==2.0.2"    # + Qdrant
pip install "hy-memory[faiss]==2.0.2"     # + FAISS
pip install "hy-memory[graph]==2.0.2"     # + Neo4j
pip install "hy-memory[mysql]==2.0.2"     # + MySQL cache
pip install "hy-memory[rerank]==2.0.2"    # + local multilingual cross-encoder rerank
pip install "hy-memory[all]==2.0.2"       # everything

2. Configuration

Only an LLM and an embedding endpoint are needed; any OpenAI-compatible service works:

export MEMORY_LLM_API_KEY="sk-your-key"
export MEMORY_LLM_BASE_URL="https://api.deepseek.com"
export MEMORY_LLM_MODEL="deepseek-chat"

export MEMORY_EMBEDDER_API_KEY="sk-your-key"
export MEMORY_EMBEDDER_BASE_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
export MEMORY_EMBEDDER_MODEL="text-embedding-v3"
export MEMORY_EMBEDDING_DIMS=1024

Or use the OpenAI defaults with a single key:

export OPENAI_API_KEY="sk-your-key"

See the Environment Variable Reference for the full list.

3. Chat Memory

from hy_memory import HyMemoryClient

client = HyMemoryClient(mode="pro")

# Write: plain text or OpenAI-format messages
client.add("I love sci-fi movies, especially Interstellar", user_id="user_1")
client.add([
    {"role": "user", "content": "Recommend a movie"},
    {"role": "assistant", "content": "Try Interstellar β€” a sci-fi masterpiece by Nolan"},
], user_id="user_1")

# Search: returns the profile / normal / proactive channels
results = client.search("What movies does the user like?", user_ids=["user_1"], limit=10)
for mem in results["memories"]["normal"]:
    print(f"  [{mem['score']:.2f}] {mem['content']}")

client.close()

4. Working & Coding Memory

Writes must include a tool-call trace (tool_calls or role="tool" messages) and take effect in pro / ultra mode only (lite ignores scene and stores everything as raw memory):

messages = [
    {"role": "user", "content": "Fix the timeout in the search client."},
    {"role": "assistant", "content": "", "tool_calls": [{
        "id": "call_1", "type": "function",
        "function": {"name": "apply_patch", "arguments": '{"path":"src/search.py"}'},
    }]},
    {"role": "tool", "tool_call_id": "call_1", "name": "apply_patch",
     "content": "Patch applied successfully"},
]

client.add(
    messages,
    scene="coding",                              # or "working"
    user_id="user_123",
    workspace_id="github.com/example/service",   # optional: isolate by repository / project
    branch="fix/search-timeout",                 # optional: requires workspace_id
)

result = client.search(
    "Which timeout implementation was finally kept?",
    scene="coding",
    user_ids=["user_123"],
    workspace_id="github.com/example/service",
    branch="fix/search-timeout",
    limit=5,
)

result["facts"]          # fact layer
result["experiences"]    # experience layer
result["chain_block"]    # coding only: version-evolution context

5. Routing, Source Isolation & Advanced APIs

  • Automatic scene routing: When scene is omitted, writes without tool calls go straight to chat (zero LLM calls), and writes with tool calls are classified once into chat / working / coding. For searches, search_mode="default" goes to chat with zero LLM calls, while search_mode="agentic" performs scene classification, query rewriting, and time-range parsing. See Working/Coding Scenes.
  • Separate memories by source: Pass source="openclaw" | "cursor" | "codex" on add() and filter with include_sources / exclude_sources on search() to isolate caller applications independently of agent_id.
  • CRUD & System 2 evolution: client.get(id), client.update(id, text), client.delete(id), client.list_memories(user_id=...), and client.evolve(user_id=...) (triggers System 2 in ultra mode).
  • Per-call embedding endpoint override: Pass embed_override={"base_url": ..., "api_key": ...} to add() / search() or use with embed_endpoint_override(...): for concurrency-safe endpoint switching.

For provider configuration, storage backends, scenes, and the complete API, see the Usage Guide.

System Architecture

Hy Memory architecture

Hy Memory provides three memory scenes (scene), one per kind of task. add() / search() go through the scene router first, then into the Chat or the Working / Coding pipeline (orange parts in the diagram are coding-only; without them the pipeline is Working):

sceneTasksMemory shape
chatSmall talk, personal information, preferences, general Q&AHierarchical memory (L0–L7)
workingProductivity work: documents, spreadsheets, research, financial analysis, content creation, computer operationTwo-layer memory: session facts + reusable experiences
codingSoftware engineering: coding, debugging, testing, builds, APIs/SDKs, scripts, SQLThe same two-layer memory as working, plus a code version-evolution chain (diff chain)

1. Chat: Memory Hierarchy

For long-term personal memory. Raw conversations are abstracted layer by layer into facts, summaries, a profile, and higher-level cognitive structures:

LayerNameContent
L0BASIC_INFOBasic information
L1RAWRaw conversation (append-only)
L2FACTAtomic facts (versioned, immutable records)
L3SUMMARYSession summaries
L4IDENTITYReserved
L5KNOWLEDGEReserved
L6SCHEMAMental models (abstract narrative templates, graph layer)
L7INTENTIONProspective intentions (with an expiry; demoted to ordinary facts once expired)

Modes

ModeWhat it doesGraph DBUse case
liteRaw-only embedding writes and searches, no LLM calls; scene and layers are ignoredNoVanilla vector RAG, fast ingestion, zero LLM cost
pro+ LLM extraction + reconcileOptional (MEMORY_ENABLE_GRAPH, off by default)Standard workloads
ultra+ System 2 schema induction + cross-domain sweeper (triggered by evolve())Yes (always on)Full cognitive architecture

2. Working: Two-Layer Productivity Memory

For getting things done. One session with tool calls (a trace) produces both layers at once:

LayerWhat it recordsPer sessionStructure
Facts (WorkingFact)What this session did: the task, the approach, the resultExactly 1, no thresholdstatement, subject, files, details (verbatim source)
ExperiencesApproaches that can be reused next time a similar task comes up0–N, must pass a value thresholdtask, search_keys, solution, boundary_envs, confidence

The fact layer guarantees that work which was done is remembered; the experience layer makes what was learned reusable. The two layers link to each other (fact.derived_experience_ids ↔ experience.source_fact_ids), and superseded facts stay traceable through a task-progress chain (fact_chain_summary).

3. Coding: Two-Layer Memory + Version-Evolution Chain

Coding reuses the same two-layer pipeline and data schema as Working. It differs in three ways:

WorkingCoding
Extraction specProductivity-task specSoftware-engineering spec: the subject must be code or a decision; record exact coordinates (file / function / branch) instead of pasting code
Storagecoding_memory.db + *_coding_keys / *_working_facts collectionscoding_memory_cv.db + *_cv collections, fully isolated from working
Version-evolution chainNoneYes: how many times the same code was changed, and what each change was

Why a version-evolution chain: fact extraction deliberately avoids pasting code (the budget goes to more informative coordinates), so a version that was later reverted never appears in a fact. That version is usually not in the working tree or at HEAD either, so it cannot be read back afterwards. The chain stores the real diff of every version with zero LLM calls, filling in exactly the information a summary inevitably drops.

Design highlights (hy_memory/coding/chain_store.py):

  • A decorator, not a retriever: the vector layer decides which facts are returned; the chain only attaches, under each fact, every revision of the code it touches (the user’s words for that revision + the real diff).
  • Change sources: local Git diffs (set MEMORY_CODING_CHAIN_REPO_ROOT), or deterministic replace actions with path / old / new in the tool trace.
  • Chain granularity & isolation: one thread corresponds to one block of one function in one file on one branch, isolated by user_id + workspace_id.

Support

Storage Backends

  • Vector stores: ChromaDB (default), Qdrant, FAISS
  • Graph stores (ultra mode): Kuzu (default), Neo4j
  • Working / Coding metadata: local SQLite

Documentation

Graph View

Backlinks