TencentDB Agent Memory is a TypeScript-based system designed to provide AI agents with fully local, long-term memory capabilities without requiring external API dependencies. The project implements a sophisticated 4-tier progressive pipeline architecture that fundamentally rethinks how agents store and retrieve information across sessions and tasks.
The core innovation of TencentDB Agent Memory lies in its rejection of flat storage models in favor of a layered, hierarchical approach combined with symbolic memory representation. Rather than indiscriminately accumulating conversation history or performing irreversible lossy summarization, the system organizes memory into distinct layers with different storage mechanisms and retrieval patterns. Short-term context is layered from raw tool outputs at the bottom through step-level summaries in the middle to lightweight Mermaid canvas representations at the top. Long-term personalization follows a semantic pyramid structure progressing from raw conversations at L0 through atomic facts at L1, scenario blocks at L2, and user profiles at L3. This layering enables progressive disclosure where agents attend to high-level structures in context and drill down to lower layers only when specific details are needed.
The symbolic memory component addresses the token consumption problem inherent in long-running agent tasks. Rather than including verbose intermediate logs, search results, code snippets, and error traces directly in context, the system encodes task state transitions in Mermaid syntax—a format that is both precise enough for language models to parse and concise enough for human inspection. Full tool logs are offloaded to external files while only a lightweight task map remains in the active context. Agents can trace back to complete raw text through node_id references, maintaining full traceability while dramatically reducing token usage.
The repository demonstrates substantial performance improvements when integrated with OpenClaw. On the WideSearch benchmark, the system increases success rate from 33% to 50% while simultaneously reducing token usage by 61.38%. On SWE-bench, which runs 50 consecutive tasks per session to simulate real-world context accumulation pressure, success rate improves by 9.93% while token usage drops by 33.09%. For long-term memory capabilities, PersonaMem accuracy increases from 48% to 76%. These benchmarks are measured over continuous long-horizon sessions rather than isolated turns, reflecting the system's design for sustained agent operation.
The implementation supports multiple agent frameworks. For OpenClaw integration, the system operates as a plugin with zero-configuration setup, defaulting to a local SQLite backend with sqlite-vec for vector operations. Optional short-term compression features are available for versions 0.3.4 and above. The system also supports Hermes agents through both Docker deployment for greenfield setups and integration into existing Hermes installations without containerization.
The architecture maintains full traceability and lossless recovery by preserving deterministic paths from high-level abstractions back to ground-truth evidence. Whether dealing with offloaded error logs or distilled user preferences, the system guarantees complete drill-down chains from top-layer symbols through mid-layer indices to bottom-layer raw text. This design philosophy ensures that compression never sacrifices the ability to verify or recover original information.
The project is classified within the agent, AI-agent, embedding, LLM, local-first, long-term-memory, memory, and openclaw-plugin domains, reflecting its positioning as infrastructure for persistent agent cognition. The dual-layer storage strategy persists facts, logs, and traces in databases for robust full-text retrieval while storing personas, scenes, and canvases as human-readable Markdown files for high information density and transparent inspection.
