Does this require an internet connection or cloud service?

No. It's designed as a local-first system running on your own machine via three local services (memory-core, memory-hub, proxy); you supply your own LLM API parameters in the .env file.

Which agent frameworks does it work with?

The README lists native support/config docs for DeepSeek Harness, Claude Code, Codex, CodeBuddy, WorkBuddy, Hermes, and OpenClaw, with a generic integration guide for other frameworks via the Proxy.

Do I need to install a plugin or MCP server in my agent?

No. You point the agent's base URL at the Proxy and it uses the same protocol it already speaks; no plugin, hook, or MCP server is required.

The repository badge indicates MIT, though the metadata field on this listing shows unknown; check the LICENSE file in the repo for confirmation.

How is this different from standard RAG over documents?

Standard RAG answers what can be found; this system also tracks who can use an asset, which version is valid, and which agent it's assigned to, via layered memory (L0-L3), Skills with versioning, Wiki link graphs, and CodeGraph impact analysis.

Can I migrate data from an older version?

Yes, a migration tool is included at MemoryCore/scripts/migrate-v2-to-v3 for moving data from v1.x/v0.x installs to v2.0.0+; new installations don't need it.

TencentDB Agent Memory: Local-first long-term

TencentDB Agent Memory is a self-hosted memory system for AI agents that stores conversations, distilled facts, reusable skills, document knowledge, and code structure in one place. It's built for developers and teams running multiple agents (Claude Code, Codex, CodeBuddy, OpenClaw, Hermes, and others) who want those agents to stop re-learning the same project context in every session.

The core idea: instead of treating memory as a chat log dump, the system organizes information into distinct asset types (Chat Memory, Skills, Wiki, CodeGraph), each with ownership, versioning, and access control. Any agent connected through the proxy can read from and write to the same memory pool, subject to visibility rules (private, team, restricted, agent). This lets a team of agents share proven workflows and project knowledge without leaking everything to everyone.

Key features

Layered memory pipeline (L0-L3): raw conversations are distilled step by step into atomic facts, scenario summaries, and long-term personas, so retrieval can pull a quick context bootstrap or drill down to exact wording when needed.
Chat Memory: retains user preferences, facts, decisions, and interaction history per agent, so you don't have to re-explain context every session.
Skill library: captures reusable, versioned workflows (troubleshooting steps, review checklists) extracted from past work, with trigger boundaries and validation rules, not just prompt snippets.
Wiki and CodeGraph: turns documents into linked, structured pages and indexes code symbols, call relationships, and impact paths so agents can check what a change might affect before editing.
Zero-code agent integration: a single Proxy speaks the same protocol agents already use, so pointing an agent's base URL at the Proxy is enough, no plugin or MCP server required.
Access control model: assets are owned, versioned, and shared explicitly (private by default) via team, role, or per-agent ACLs, with a System Admin and Team Admin role split.
Cold-start import: existing codebases, documents, and past agent conversation sessions can be imported and auto-processed into Wiki, CodeGraph, Skills, and Chat Memory.
Local-first deployment: runs as three local services (memory-core, memory-hub, proxy) with a web panel for managing teams, agents, and memory assets.

Ideal use cases

Running several coding or ops agents (Claude Code, Codex, CodeBuddy, etc.) against the same codebase or project, where you want shared context instead of re-explaining the setup to each one.
Building a small "agent team" (researcher, builder, reviewer) that needs different but overlapping memory loadouts, with explicit control over what each agent can see.
Onboarding a new agent or teammate onto an existing project by importing prior conversations, docs, and code instead of starting from a blank context window.
Teams that need auditability over what an agent "knows": ownership, version, status, and usage counts on each memory asset.
Not a good fit if you just need simple session-scoped memory for a single chatbot, or if you want a hosted/managed vector database with no self-hosting overhead. It's also not a drop-in RAG replacement for pure document search use cases that don't need ownership, ACLs, or agent-specific loadouts.

Installation

The project ships a one-command setup that starts all three services (memory-core, memory-hub, proxy):

git clone https://github.com/Tencent/TencentDB-Agent-Memory.git
cd TencentDB-Agent-Memory/deploy/global-images
cp .env.example .env
$EDITOR .env       # fill in LLM parameters for the memory group and proxy group
./start-all.sh     # launches everything; prints a one-liner to paste into Claude when done

Open the management panel at http://localhost:8125.

Requirements include Node.js 22.16+ and, for OpenClaw integration, OpenClaw 2026.3.13+. Full setup docs (standalone Memory Hub deployment, Proxy configuration for Claude Code / CodeBuddy, stopping and cleanup, port reference) live in INSTALL.md. If you're upgrading from an older v1.x/v0.x install, a data migration tool is provided under MemoryCore/scripts/migrate-v2-to-v3.