Deploy TencentDB Agent Memory

Deploy and Host TencentDB Agent Memory on Railway

> ## 在 Railway 上一键部署 TencentDB Agent Memory —— 让 Claude Code、CodeBuddy、DeepSeek、Codex 等编码智能体拥有可自托管的团队共享记忆 > 三个服务(memory-core · memory-hub · memory-proxy)一键上线,把对话、文档和代码沉淀为可复用的团队记忆资产。

Give Claude Code, Codex, DeepSeek CodeBuddy (and many others) a shared, self-hosted team memory — point their base URL at your own proxy and every agent remembers.

TencentDB Agent Memory is an open-source AI agent memory hub (25k+ GitHub stars) that turns conversations, docs, and code into four reusable memory assets — Chat Memory, Skills, LLM-Wiki, and Code-Graph — shared across agents, frameworks, and teammates. This Railway template deploys the full three-service stack (memory core, memory hub panel, and LLM proxy) in one click: fill in three LLM fields and you have persistent, self-hosted LLM memory for your whole team's coding agents.

About Hosting TencentDB Agent Memory on Railway

Self hosting agent memory normally means wiring three services together: a private memory engine, a web panel + knowledge service, and an OpenAI/Anthropic-compatible proxy that injects memories into every request. This template ships all three as pinned Docker images with Railway-ready defaults baked in — private networking between services, volumes for persistence (SQLite, no external database needed), healthchecks, and secure-by-default networking (the memory core is never exposed publicly; the knowledge service stays internal unless you opt in). Cross-service credentials and LLM settings are prewired with reference variables: enter your LLM endpoint, key, and model once, and everything inherits.

Why Deploy TencentDB Agent Memory on Railway

• One-click, three services — no Docker Compose, no VPS setup, no YAML mounting. The upstream images only read config from mounted files; this template's wrappers render that config from environment variables automatically.
• Your keys, your data — memories, skills, and wiki content live in SQLite on your own Railway volumes. Nothing routes through a third-party memory SaaS.
• Usage-based cost — a light team stack typically idles in the single digits per month; you pay for what you use rather than per-seat memory pricing. New Railway users get a free trial to test it out.
• Battle-tested defaults — heap caps to prevent silent OOM kills on small plans, private-network wiring, and an included end-to-end smoke test (test/smoke-test.sh) that verifies chat round-trip, capture, and recall.

Common Use Cases

• Shared memory for Claude Code — stop re-explaining your architecture every session; project context, decisions, and constraints auto-inject into every request.
• Team knowledge that compounds — onboard a new agent (or teammate) with the team's accumulated Chat Memory, Skills, wiki, and code graph on day one.
• Multi-agent teams — equip a Scout, Builder, and Reviewer agent with different memory loadouts from one control panel.
• Cold-start from existing assets — import codebases, docs, and past agent sessions; Code-Graph and LLM-Wiki index them automatically.

Dependencies for TencentDB Agent Memory Hosting

None outside this template — no external database. Three services deploy together:

Implementation Details

• Official multi-arch images (agentmemory/*), version-pinned for deliberate updates
• Fill three variables at deploy: MEMORY_LLM_BASE_URL, MEMORY_LLM_API_KEY, MEMORY_LLM_MODEL (any OpenAI-compatible endpoint — OpenRouter, DeepSeek, etc.); a PROXY_UPSTREAM_* group lets you route user chats to a stronger model than memory jobs
• Admin key auto-generated; Panel is key-gated; memory layers L0 (chat logs) → L1 (facts) → L2 (scenes) → L3 (profile) extract asynchronously
• Verify any deployment end-to-end with the bundled smoke test

How to Use TencentDB Agent Memory (after deploy)

1. Open the memory-hub domain and log in with your ADMIN_USER_KEY. You will find this by clicking into memory-core > Variables in the Railway project.
2. Create a Team → Agent → Task in the Panel.
3. Point your coding agent at the proxy:

export ANTHROPIC_BASE_URL=https:///claude-code/default
export ANTHROPIC_AUTH_TOKEN=''
claude --model 

4. Chat normally. The proxy injects that agent's memories into every request and captures conversations back into team memory. Supported clients include Claude Code, Codex, CodeBuddy, WorkBuddy, OpenRouter, and DeepSeek Harness — zero plugins, hooks, or MCP servers required.

Full docs here

How TencentDB Agent Memory Compares

vs. plain RAG / vector databases — RAG answers "what can be found?"; Agent Memory also answers who can use it, which version is valid, and which agent should receive it: ownership, versions, visibility (private/team/ACL), and per-agent loadouts. Upstream's PersonaMem benchmark shows +59% relative improvement in long-horizon user understanding.

vs. Mem0 / Zep (memory APIs) — those are developer SDKs you integrate into your own app code. Agent Memory needs zero code changes: it sits in front of your existing coding agents as a proxy and manages memory for a whole team, with a human-controlled review panel.

vs. ChatGPT / Claude built-in memory — vendor memory is personal, opaque, and locked to one product. Here memory is a governed team asset that moves across frameworks — and it's yours, on your infrastructure.

vs. chat history dumps — raw logs don't transfer experience. Layered extraction distills conversations into atomic facts, scenarios, and profiles, plus executable Skills with versions and trigger rules.

Frequently Asked Questions (FAQs)

What is TencentDB Agent Memory? An open-source, self-hosted memory hub for AI agents from TencentCloud — shared Chat Memory, Skills, LLM-Wiki, and Code-Graph for coding agents like Claude Code, with a web panel for human review and access control.

Is it free? Yes — the upstream project is open source and this packaging is MIT-licensed. You only pay Railway infrastructure costs, plus your own LLM API usage.

Which LLMs work? Any OpenAI-compatible endpoint (OpenRouter, DeepSeek, etc.); Anthropic protocol is also supported. You can use a cheap model for memory processing and a strong one for actual coding chats.

Does my data persist across redeploys? Yes — memories, skills, and wiki content live on Railway volumes (SQLite). Redeploys and image updates keep your data.

Is it secure to expose? The memory core never gets a public domain (enforced by the template), the panel requires a key on every visit, and the knowledge service stays private unless you explicitly publish it. Mark your LLM keys sealed for write-only storage.

How do I verify my deployment works? Run the included test/smoke-test.sh — it checks health, sends chats through the proxy, and confirms a brand-new session can recall stored facts.