Comparison of TencentDB's agent memory, which excels at compressing messy run histories for token savings, versus the Memos local plugin, which focuses on turning execution history into reusable habits and long-term learning through feedback loops.
tencentdb agent memory getting open-sourced made me rethink agent memory a bit. what i like most is its short-term context cleanup. agent runs get messy fast: tool logs, retries, failed branches, repeated observations, and a lot of stuff you probably don’t want dumped back into the prompt. tencentdb’s mermaid-style canvas feels practical here. it compresses a messy run into something easier to inspect, while node\_id still lets you trace back to the raw data. the claimed token saving, up to 61.38%, is also meaningful if you are running agents on real tasks. i also like that it is not just one giant vector db. conversation records, atomic facts, scenario memory, and profile memory are separated, with sqlite / sqlite-vec and markdown files keeping things fairly local and inspectable. so yeah, tencentdb looks strong for short-term memory management. but compression is not the same thing as learning. if an agent spends an hour debugging docker permissions and finally finds a uid/gid mismatch, i don’t just want a cleaner summary of that run. i want the agent to check uid/gid earlier next time and stop starting with chmod 777. that is not just shorter memory. that is a reusable debugging habit. this is where memos local plugin 2.0 feels like it is solving a different layer of the problem. its focus seems less about reducing token cost but more about turning execution history into better future behavior. that’s a different view. the trace layer keeps the step-level record. the policy layer distills patterns across tasks. the world model stores environment-level knowledge. then useful repeated patterns can become reusable skills. that feels closer to long-term agent learning than long-term storage. the feedback loop is the part i care about most. if a task fails, i don’t want the system to neatly save that failure and accidentally retrieve the same bad path next week. i want the failed path to become less likely. step-level feedback, task-level feedback, llm scoring, and reward propagation all sound like attempts to make memory actually change future decisions. the observability side matters too. tencentdb’s markdown-inspectable memory is nice, but the local plugin having a vite viewer ui, live event stream, and structured logs feels more useful when you are trying to understand why an agent picked a certain policy or skill. so i don’t really see tencentdb and memos local plugin as direct competitors. tencentdb seems very strong at making memory manageable: compress the messy run, reduce token cost, keep it inspectable, and preserve traceability through node\_id in a short-term way. but the local plugin feels more like the long-term answer. it is less about storing or compressing what happened, and more about turning traces, feedback, and repeated patterns into better future behavior. to me, tencentdb answers: “how do we manage what just happened?” memos answers: “how do we make the agent stop making the same mistake again?”
TencentDB Agent Memory is an open-source tool providing symbolic short-term and layered long-term memory for AI agents, reducing token usage by up to 61.38% and improving task success rates by over 50%.
Tencent Cloud database team open-sourced TencentDB Agent Memory, a runtime system that solves the context degradation problem in long tasks for AI agents, compressing short-term context into the memory system through three-layer backtracking and dynamic compression, and integrating a long-term memory pipeline. This is a landmark attempt for AI agent memory systems moving from 'database' to 'runtime'.
Tencent open-sourced TencentDB-Agent Memory, a symbolic short-term and layered long-term memory system for AI agents, which cuts token usage by up to 61% and improves persona accuracy from 48% to 76%.
Tencent has open-sourced TencentDB Agent Memory, which solves the AI agent long-context overflow problem through hierarchical memory management (symbolic short-term memory + hierarchical long-term memory). Benchmarks show token consumption reduced by up to 61% and task success rate improved by over 50%.