Tencent has released TencentDB Agent Memory, an open-source memory system for AI agents under the MIT license. The project addresses the critical challenges of context bloat and recall failure common in long-horizon agent deployments.
The architecture combines symbolic short-term memory with layered long-term memory. It integrates natively with the OpenClaw framework as a plugin and supports the Hermes Agent via a Gateway adapter. To ensure privacy and performance, the system uses a local SQLite backend with the sqlite-vec extension, eliminating the need for external API calls for memory retrieval.
Instead of the standard flat vector store approach, TencentDB Agent Memory implements a 4-tier semantic pyramid: L0 Conversation (raw logs), L1 Atom (atomic facts), L2 Scenario (scene blocks), and L3 Persona (user profiles). The system prioritizes the Persona layer for daily user preferences, drilling down to Atoms or raw Conversations only when fine-grained evidence is required. Storage is heterogeneous, with structured data in DBs for full-text search and high-level artifacts like Personas stored as human-readable Markdown files.
For short-term memory management, the system utilizes Mermaid syntax to create a symbolic representation of task states. Verbose tool logs and search results are offloaded to external reference files. The agent reasons over a lightweight symbolic graph in its context window and only retrieves raw text via specific node IDs when necessary. This deterministic drill-down approach minimizes token overhead during complex reasoning tasks.
Benchmark results indicate significant performance gains. On WideSearch, the plugin increased the pass rate from 33% to 50% while reducing token usage by 61.38%. On SWE-bench, success rates climbed from 58.4% to 64.2%, with a 33.09% reduction in token consumption. Accuracy for long-term persona recall also saw a jump from 48% to 76%.
