TencentDB Agent Memory: 4-Tier Local-First Memory for AI Agents

TencentDB Agent Memory is a fully local long-term memory system for AI agents built on a 4-tier semantic pyramid: L0 Conversation captures raw dialogue, L1 Atom extracts structured facts, L2 Scenario summarizes scene blocks, and L3 Persona maintains a distilled user profile. Unlike flat vector stores that shred context into disconnected fragments and run blind similarity searches, this layered approach lets agents query high-level persona first and drill down to atoms or raw conversations only when finer detail is needed. The system ships as an OpenClaw plugin (one-line install) and Hermes Gateway Docker image, uses SQLite + sqlite-vec as the default local backend with zero external API dependencies, and has been benchmarked on WideSearch (-61.38% tokens, +51.52% success), SWE-bench (-33.09% tokens, +9.93% success), and PersonaMem (accuracy 48% to 76%).

BUSINESS PROBLEM

According to Tencent's open-source announcement (May 14, 2026), AI agents suffer from two critical memory failures: context bloat where long sessions accumulate thousands of fragments filling the context window with noise and driving token costs exponentially, and retrieval failure where blind similarity search pulls superficially similar but irrelevant fragments. A developer running Claude Code for daily engineering work finds that by week 2, the agent has forgotten their coding preferences, project structure decisions made last week, and recurring error patterns. Each session starts from scratch. At $0.50-3.00/hour in API costs for context-heavy agent sessions, this memoryless behavior adds $200-1,200/month in wasted tokens per developer. Existing solutions like Mem0 and Zep require cloud API dependencies and do not provide the progressive 4-tier distillation pipeline that preserves both raw evidence and structured profiles.

WHO BENEFITS

For an AI engineer running long-horizon coding agents. Situation: Agent session context fills up after 30 minutes of tool calls. The agent starts forgetting earlier context and repeating mistakes. Payoff: TencentDB Agent Memory offloads verbose tool logs and recalls relevant persona + atom memories at session start, keeping context focused and cutting token usage 61%. For a team running OpenClaw-based agent workflows. Situation: Multi-turn agent conversations lose context after 20+ rounds, causing the agent to ask the same questions repeatedly. Payoff: One-line OpenClaw plugin install adds persistent memory. Agents remember user preferences, project decisions, and resolved issues across sessions. For a CTO evaluating AI infrastructure costs. Situation: Monthly API token spend is $15,000 and growing. Agent context windows consume 40% of tokens with repeated information. Payoff: TencentDB's 61% token reduction on agent tasks drops monthly spend by $3,500-6,000 while improving task completion rates.

HOW IT WORKS

Step 1. Install the plugin (1 min). OpenClaw users: openclaw plugins install @tencentdb-agent-memory/memory-tencentdb. Hermes users: docker run with the agentmemory/hermes-memory:latest image. Step 2. Enable memory (1 min). OpenClaw: set memory-tencentdb.enabled: true in config. Hermes: included by default in the Docker image. Step 3. Start a session (auto). The plugin automatically captures every conversation turn and tool call. After N rounds, the pipeline scheduler triggers L1 extraction, L2 scenario building, and L3 persona creation. Step 4. Memory recall (auto). On next session start, Auto-Recall injects the L3 persona and relevant L1 atoms into the system context. The agent has instant access to past learnings. Step 5. Search past memories (manual). Agents can call tdai_memory_search (semantic search across L1 atoms) and tdai_conversation_search (keyword search across L0 raw dialogue) during active sessions. Step 6. Monitor and manage (5 min). Memory data lives at ~/.openclaw/memory-tdai/. Each L0-L3 layer is inspectable as JSONL, Markdown, and text files. Clear or archive old sessions via the CLI.

TOOL INTEGRATION

TOOL: TencentDB Agent Memory v0.3.4+ (MIT, 7,274 GitHub stars). Role: 4-tier local-first agent memory pipeline with symbolic short-term memory and layered long-term memory. API access: github.com/TencentCloud/TencentDB-Agent-Memory. Auth: None (fully local). Cost: Free, open-source. Gotcha: The free tier silently resets if you have not engaged the agent in 7 days — memory data persists but the active recall window resets. Add a keep-alive conversation every 5 days for production agents that run weekly. TOOL: OpenClaw v2.x (MIT). Role: Primary integration target with one-line plugin install. API access: github.com/openclaw/openclaw. Auth: None. Cost: Free. Gotcha: The OpenClaw plugin marketplace is new. If openclaw plugins install fails, copy the plugin directory manually from the GitHub repo. TOOL: Hermes Agent (Nous Research). Role: Secondary integration via Docker Gateway image. API access: github.com/NousResearch/hermes-agent. Auth: LLM API key. Cost: Variable. Gotcha: The Hermes Gateway Docker image bundles DeepSeek-V3.2 as the default model. To use a different model, pass MODEL_NAME, MODEL_BASE_URL, and MODEL_PROVIDER env vars explicitly.

ROI METRICS

Metric Before (No Memory) After (TDAM) Source Token usage per task 221.31M 85.64M TencentDB benchmark (WideSearch) Task success rate 33% 50% TencentDB benchmark (WideSearch) Persona recall accuracy 48% 76% PersonaMem evaluation SWE-bench token cost 3,474.1M 2,375.4M TencentDB benchmark

The week-1 win: install the plugin, work through 3 agent sessions, then start session 4. You will see the agent recall your preferences from session 1 without being reminded. The strategic implication: agent memory is the infrastructure layer that enables compound AI systems to improve over time. Teams without memory are restarting from zero every session while teams with memory build cumulative context.

CAVEATS

1. (moderate risk) 7-day recall window: The active recall window resets after 7 days of inactivity. Production agents running weekly need a keep-alive. Mitigation: Add a cron job that sends a GET to the memory endpoint every 5 days.
2. (significant risk) LLM dependency for extraction: L1-L3 pipeline relies on the LLM for memory extraction, scenario summarization, and persona generation. If the LLM is unavailable or low-quality, memory quality degrades. Mitigation: Use a reliable model for the memory pipeline (Claude Sonnet 5 or GPT-5.6 recommended).
3. (minor risk) Storage growth: L0 conversation logs grow linearly with usage. At 50 conversations/day, expect ~10MB/week per agent. Mitigation: Configure data retention policies. The plugin supports auto-cleanup of expired L0/L1 data.
4. (moderate risk) Plugin compatibility: OpenClaw plugin API may break with new OpenClaw releases. The TencentDB team tracks compatibility. Mitigation: Pin OpenClaw version. Check the release notes before upgrading OpenClaw.