TencentCloud/TencentDB-Agent-Memory

TencentDB Agent Memory is a team-level memory hub for AI Agents — turning conversations, docs, and code into four reusable memory assets (Chat Memory, Skill, LLM-Wiki, Code-Graph) that are governed, shared, and equipped across agents and frameworks.

TencentDB Agent Memory

What it is – An open‑source “memory hub” that lets large‑language‑model (LLM) agents store, organise and re‑use the knowledge they generate during a workflow. It turns raw chat logs, extracted skills, documentation and code‑bases into structured assets (Chat Memory, Skills, Wiki, CodeGraph) that can be shared across many agents and teams.

Core idea

When you work with an LLM‑powered agent, the conversation, the code it writes, and the docs it reads are usually discarded after the session ends. TencentDB Agent Memory captures that experience, distils it into reusable pieces, and makes those pieces available to future agents so they don’t have to “reinvent the wheel”.

Main components

Component What it stores How it is used
Memory Core Raw conversations (L0) → extracted facts (L1) → scenario blocks (L2) → long‑term persona (L3) Retrieval is layered: quick L2/L3 context first, fall‑back to L1/L0 with BM25 + vector search when needed.
Memory Hub (Panel) Central UI for teams: create teams, users, agents; browse assets; set ACLs; bind assets to agents. Human‑in‑the‑loop control – decide what is private, team‑wide, or restricted.
Proxy Thin HTTP gateway that presents a uniform API to any LLM‑agent framework (Claude Code, DeepSeek Harness, CodeBuddy, OpenClaw, etc.). Agents only need to point their base URL to the proxy; no SDK changes required.
Wiki Structured pages generated from product docs, design specs, runbooks; linked together as a graph. Agents can call /v3/tools/list → /v3/tools/call to read a page or follow links.
CodeGraph Index of symbols, files, call relationships, impact paths for a code repository. Enables impact analysis and precise code‑search before a modification.
Skill library Versioned, validated “experience blocks” (e.g., release checklist, troubleshooting flow) extracted from past runs. Agents can load a Skill into their context on demand.

How you get it running

  1. Clone the repo and run the provided start-all.sh script in deploy/global-images. The script launches three Docker containers: memory‑core, memory‑hub, and proxy.
  2. Fill in the .env file with two sets of LLM credentials (one for the memory‑group and one for the proxy‑group).
  3. After the containers start, open http://localhost:8125 – the Memory Hub UI where you create teams, import assets, and bind them to agents.
  4. Point any supported agent framework to the proxy URL (the README lists ready‑made adapters for Claude Code, DeepSeek Harness, Codex, CodeBuddy, WorkBuddy, Hermes, OpenClaw, etc.). No code changes are required beyond the base URL.

The storage backend is PostgreSQL by default; an experimental MongoDB backend can be enabled via the install guide.

Typical workflow (cold‑start to production)

  1. Import existing knowledge – upload a code repository, docs, or past conversation logs. The hub automatically builds a CodeGraph and Wiki, and extracts Chat Memory / Skills.
  2. Create a team – add human users and virtual agents (Scout, Builder, Reviewer, …). Assign each agent the assets it needs (e.g., Builder gets the product Wiki and CodeGraph; Scout gets market‑research Wiki).
  3. Run a task – an agent calls the proxy, which pulls the appropriate assets on‑demand, performs its work, and writes back new assets (new Skills, updated Wiki, etc.).
  4. Iterate – subsequent agents start with the enriched memory, requiring fewer prompts and fewer LLM calls.

Benefits shown in the repo’s benchmark

Metric Without memory With memory Improvement
PersonaMem (ability to retain user info across long interactions) 48 % 76 % +59 %

The test demonstrates that agents equipped with the hub retain user preferences and facts far better than a naïve chat‑only approach.

Who might use this?

  • Solo developers building AI‑augmented tools who want their LLM assistants to remember past decisions.
  • Start‑ups that run multiple specialized agents (research, coding, QA) and need a shared knowledge base without re‑training models.
  • Enterprises that want fine‑grained access control (private / team / ACL) over what each agent can see.
  • Researchers exploring retrieval‑augmented generation (RAG) pipelines that combine conversation memory with code and doc indexing.

Roadmap (as of the README)

  • v2.0.1 – zero‑config cold start, faster Wiki generation, custom prompts per user/team, Skill export, Codex “IDE Plan” mode support.
  • Ongoing work on fully automated memory routing, private‑repo CodeGraph indexing, and richer ACL models.

Community & contribution

  • Open‑source under the MIT license.
  • Issues are responded to within 24 h; discussions and PRs are encouraged.
  • The project re‑uses code from the open‑source CodeGraph, Hermes Agent, and Karpathy’s “LLM Wiki” concepts, and credits those upstream projects.

TL;DR – TencentDB Agent Memory is a self‑hosted platform that gives LLM agents a persistent, searchable, and permission‑aware “brain”. By converting chats, docs, and code into layered assets, it lets teams of agents work faster, with fewer repeated prompts, and with better continuity across sessions.

Related

  • Project
    Dicklesworthstone/cass_memory_system

    cass‑memory is a CLI that aggregates session logs from many AI coding agents, extracts structured rules (playbook) with confidence tracking, and provides JSON‑based queries so any agent can retrieve relevant knowledge before coding.

  • Project
    marciopuga/cog

    A centralized, plain-text memory layer for AI agents that uses markdown files to share context and persistent knowledge across different tools and projects.

  • Project
    beilusaiying/always-accompany

    A multi-purpose AI agent framework focused on context and attention mechanisms, providing layered memory and dynamic retrieval for long-term companionship, coding, and work.

  • Project
    mem9-ai/mem9

    mem9 provides a persistent, searchable memory service for AI coding agents, offering hosted or self‑hosted deployments, multi‑agent sharing, hybrid vector/text search, smart ingest, and plugin integrations for major agent platforms.

  • Project
    NevaMind-AI/memU

    memU is a lightweight, agent‑driven personal memory system that records LLM session logs, turns useful interactions into reusable Markdown “skills”, and makes those skills searchable across sessions, agents, and devices. It works via small host‑adapter binaries (e.g., memu‑codex, memu‑claude‑code) that hook into agents’ session logs and instruction files. The core memory logic is ~500 lines of Python, storing embeddings in SQLite, Postgres, or an in‑memory store. Installation is a one‑message setup with an API key from memU Cloud (or self‑hosted). Once installed, agents automatically retrieve relevant skills before answering, and developers can feed custom conversations into memU via `memu memorize`. Licensed under Apache‑2.0.