TencentDB Agent Memory – Layered Memory System for AI Agents

1. What Is TencentDB Agent Memory

TencentDB Agent Memory is an open-source, MIT-licensed layered memory system for AI agents from Tencent Cloud Database. Its L0–L3 progressive architecture plus context offloading and Mermaid task canvases deliver short-term compression and long-term personalized memory. In production tests it cut token usage by up to 61.38% while raising task success by 51.52%, with native plugins for OpenClaw, Hermes, and similar agent frameworks.

Technical positioning and domain: Agent infrastructure focused on memory—not generic prompt tuning or RAG alone. It targets context overflow in long tasks and missing cross-session user profiles, acting as the “memory hub” of an agent stack.

Development background: From Tencent Cloud Database engineers with deep storage/retrieval expertise, motivated by OpenClaw/AutoGPT-style limits: finite context, high token bills, no durable personalization. DB experience applied to agent memory with lightweight, auditable design.

Core value: Fixes long-horizon agents via context offloading (tool dumps → external store + structured summaries) and L0–L3 abstraction from raw dialogue to user profiles—lowering cost while improving continuity and accuracy.

Technical characteristics: Four-layer progressive memory; Mermaid task graphs preserve structure after offload; human-readable artifacts for full traceability; default local SQLite + sqlite-vec, optional Tencent Cloud VectorDB (TCVDB) hybrid retrieval.

2. Key Features

• L0–L3 progressive memory: L0 raw logs → L1 atomic facts/preferences via LLM extraction → L2 scenario clustering (coding, research, etc.) → L3 user profile. Concrete evidence plus fast high-level retrieval.

• Context offloading + Mermaid canvas: Raw tool outputs (API payloads, web pages) move off-context; a Mermaid graph captures steps, dependencies, and state—often shrinking active context to 30–50% of original size.

• White-box traceability: Markdown/JSONL/Mermaid files at every layer—drill from L3 profile down to L0 dialogue for debug and audit.

• Pluggable storage: Zero-dependency SQLite + sqlite-vec by default; TCVDB for BM25 + vector hybrid at scale.

• One-click framework plugins: Native OpenClaw/Hermes integration via a single install command—no core agent rewrites.

• Live token monitoring: Per-operation token stats with before/after panels to quantify savings.

3. How to Use

1. Requirements: OpenClaw (v0.5.0+) or Hermes; Linux (Ubuntu 20.04+), macOS 12+, or Windows 10/11. TCVDB optional with Tencent Cloud credentials.

2. Install plugin:

   openclaw plugins install @tencentdb-agent-memory/memory-tencentdb
   

3. Restart gateway:

   openclaw gateway restart
   

   Confirm with openclaw plugins list → status active.

4. Optional TCVDB config in OpenClaw YAML:

   memory:
     backend: tencentdb
     tencentdb:
       endpoint: "https://your-instance.tcvdb.tencent.com"
       api_key: "your-api-key"
   

   Restart gateway after changes.

5. Validate: Run a multi-step task (e.g., “Research latest AI agent papers and write a report”). Inspect ~/.openclaw/memory/ — l0_dialogue.jsonl, l3_profile.md, etc.

4. Pros and Cons

5. Comparison with Similar Tools

Selection advice: OpenClaw/Hermes teams with long workflows and cost pressure → TencentDB Agent Memory. Multi-tenant personalization at scale → Mem0 (+ your vector infra). Fully autonomous agents in Letta's runtime → Letta despite migration cost.

6. Editor's Review

TencentDB Agent Memory is a rare systematic design for agent memory—both short-term compression and long-term profiles, not just fact snippets. Offload + Mermaid is a clever way to keep structure while dropping raw tool blobs.

Hard numbers (61% tokens, +51% success) matter for enterprise rollouts. White-box files are a debugging superpower as agents go multi-session and multi-agent.

Gaps: framework coverage and docs depth. If adapters expand, this could become a category reference.

— Deduct for ecosystem/docs; core tech and measured impact are strong.

7. Use Cases

• Coding assistants: Remember repo conventions, patterns, and todos across days-long projects.

• Deep research: Offload web dumps; Mermaid graph preserves investigation path without refeeding raw HTML.

• Workflow orchestration: Checkpoint ETL/report pipelines with resumable Mermaid state.

• Personalized support: L3 profiles carry preferences and history across sessions.

• Long document analysis: Summarize prior sections into memory while analyzing later chapters without context breaks.

8. FAQ

Q: Other agent frameworks?
A: Native OpenClaw/Hermes only for now; generic adapters planned.

Q: SQLite limits?
A: Fine under ~100k entries per session memory; migrate to TCVDB beyond that.

Q: Inspect memory contents?
A: Files under ~/.openclaw/memory/ — L0 JSONL, L1 JSONL, L2 Markdown chunks, L3 profile Markdown.

Q: Are savings guaranteed?
A: 61.38% / +51.52% are averages from Tencent's four task categories—your mileage varies with context length and tool verbosity.

Q: Time to build L3 profile?
A: Rough shape after 5–10 turns; stabilizes around 20+. Seed manually or bootstrap from historical logs/RAG for cold start.

Q: Multi-user/multi-agent?
A: Separate storage per instance; set instance_id in config; use TCVDB as shared backend for teams.

9. Project Links

• GitHub: https://github.com/Tencent/TencentDB-Agent-Memory