# TencentDB Agent Memory Hits #1 GitHub: Fix Agent Amnesia > Source: > Published: 2026-08-05 12:11:13+00:00 Tencent’s open-source AI agent memory system, [TencentDB Agent Memory](https://github.com/TencentCloud/TencentDB-Agent-Memory), hit #1 on GitHub Trending this week, adding nearly 2,000 stars in a single day to reach 14,600 total. The tool gives AI agents persistent, layered memory across sessions and team members—running entirely on local SQLite with no cloud dependencies and no API fees. If you’re building with Claude Code, Hermes, or OpenClaw, this is the tool dominating developer discussion right now. ## The Problem: AI Agent Memory Loss Costs Your Team Every Day Bigger context windows are the wrong answer to agent amnesia. They postpone the problem—they don’t fix it. When five developers each spend time teaching their Claude Code instance the same things about their build system, that’s four wasted lessons repeated across every new session. The root issue isn’t context size. It’s that agents have no structured way to accumulate and share institutional knowledge. The performance numbers make the cost concrete. On the SWE-bench 50-consecutive-task benchmark, agents using TencentDB Memory achieved a 64.2% success rate versus 58.4% without it. Token usage dropped from 3.47 billion to 2.38 billion—a 33% reduction. On WideSearch benchmarks, the gains were larger: [51.52% better task success and 61.38% fewer tokens](https://www.marktechpost.com/2026/05/23/tencent-open-sources-tencentdb-agent-memory-a-4-tier-local-memory-pipeline-for-ai-agents/). For high-volume agentic pipelines, 61% token savings isn’t a performance metric—it’s a cost metric that changes what’s economically viable. ## How the Four-Tier Architecture Solves Agent Memory Rather than dumping everything into a flat vector store, TencentDB Agent Memory organizes memory in a four-tier pyramid: L0 (raw conversation), L1 (extracted atomic facts), L2 (recurring scenario patterns), and L3 (synthesized team persona). Agents query from the top down—persona first, raw transcript only if needed. That inverted retrieval is what makes the token savings real. The system also handles short-term context compression for long-running tasks. Instead of bloating the active context window with verbose tool logs and error traces, it offloads those to reference files and encodes state transitions as compact Mermaid diagrams. Full traceability is preserved; the context stays lean. Beyond chat memory, the system manages four asset types: Skills (reusable procedures), LLM-Wiki (structured documentation), and Code-Graph (code relationship maps)—all shared across agents and sessions. Related:[Claude Code Sub-Agent Model Routing: Cut Costs Without Cutting Quality] ## Getting Started: One Command for OpenClaw and Hermes For OpenClaw users, the install is a single line. For Hermes users, a Docker container bundles the gateway and dependencies. MIT licensed, runs on local SQLite plus sqlite-vec, no external credentials required. ``` openclaw plugins install @tencentdb-agent-memory/memory-tencentdb ``` Hermes integration is documented in the [complete setup guide at Regolo.ai](https://regolo.ai/tencentdb-agent-memory-the-complete-guide-to-persistent-memory-for-hermes-and-openclaw-with-zero-data-retention/). Watch two gotchas: the plugin directory must be named `memory_tencentdb` with an underscore (not a hyphen), and BM25 defaults to Chinese tokenization—switch to `"en"` for English deployments or recall quality will suffer. Node.js 22.16.0 or later is required. The data lives in human-readable files under `~/.openclaw/memory-tdai/` . L2 scenario files and the L3 persona document are plain Markdown. No black box, no guessing what the agent retained. ## What TencentDB Agent Memory Misses: Storage Is Not Adjudication TencentDB stores memories. It doesn’t resolve conflicts between them. As one [DEV Community analysis](https://dev.to/dennis_pilarinos/team-memory-hubs-for-ai-agents-what-tencentdb-agent-memory-solves-and-what-it-misses-16ja) put it: “A memory hub stores; it doesn’t adjudicate.” When two agents record contradictory facts about the same API endpoint, retrieval picks one arbitrarily. March documentation persists after a June refactoring with the same confidence score it carried when accurate. Three failure modes follow: Conflict (contradictory stored facts), Staleness (memories decay against evolving codebases), and Trust (access controls govern who reads memories, not whether those memories are still correct). For teams with stable, slow-changing codebases, these gaps are manageable. For teams shipping fast, they’re a real risk. Compared to hosted alternatives like Mem0 (which abstracts storage entirely) or Zep (which uses a temporal knowledge graph to handle time-aware facts), TencentDB trades sophistication for sovereignty: it’s your data, your hardware, your control—but you own the maintenance burden too. Related:[Open-Weight AI 2026: When Self-Hosting Beats the API] ## Key Takeaways - TencentDB Agent Memory hit #1 GitHub Trending today with 14,600 stars—driven by real developer demand for persistent AI agent memory, not hype - Its four-tier memory pyramid cuts token usage 33–61% across benchmarks by querying high-level persona first and raw conversation only when needed - Installation for OpenClaw is a single npm command; Hermes users get a Docker container; MIT licensed with zero cloud dependencies - The architecture stores memories but doesn’t resolve conflicts or detect staleness—fast-moving codebases should account for this before committing - For teams with data privacy requirements or high agentic token costs, this is the most credible local alternative to hosted solutions like Mem0