Tencent's agent memory that never phones home
Most of the agent-memory products that raised money this year, Supermemory, Walrus, and the rest, are services you send your data to. Tencent Cloud's open-source TencentDB-Agent-Memory, trending on GitHub today, is the opposite bet: fully local long-term memory with zero external API dependencies, running on a plain SQLite backend under an MIT license.
The design is a four-tier pipeline. Raw conversations come in at L0, get distilled into atoms at L1, grouped into scenarios at L2, and rolled up into user personas at L3. On top of that sits what they call symbolic memory, compressing verbose task logs into compact Mermaid diagrams the agent can reload, so it keeps the shape of past work without dragging the whole transcript back into context. Retrieval is hybrid, keyword plus semantic. And because every layer is a readable artifact, you can actually open it up and debug what your agent remembers, which most black-box memory services won't let you do.
The numbers they report are the reason to look: up to 61% less token usage from offloading context, and a 51% relative jump in task success rate on their benchmarks. It already plugs into OpenClaw and Hermes.
The interesting split here is rent versus run. The funded memory startups are betting you'll pay a service to remember for you. Tencent is betting a lot of teams want memory that lives on their own disk, costs nothing per call, and never leaves the building. For agents handling anything sensitive, that second bet looks strong. Repo at github.com/TencentCloud/TencentDB-Agent-Memory.
← Back to all articles
The design is a four-tier pipeline. Raw conversations come in at L0, get distilled into atoms at L1, grouped into scenarios at L2, and rolled up into user personas at L3. On top of that sits what they call symbolic memory, compressing verbose task logs into compact Mermaid diagrams the agent can reload, so it keeps the shape of past work without dragging the whole transcript back into context. Retrieval is hybrid, keyword plus semantic. And because every layer is a readable artifact, you can actually open it up and debug what your agent remembers, which most black-box memory services won't let you do.
The numbers they report are the reason to look: up to 61% less token usage from offloading context, and a 51% relative jump in task success rate on their benchmarks. It already plugs into OpenClaw and Hermes.
The interesting split here is rent versus run. The funded memory startups are betting you'll pay a service to remember for you. Tencent is betting a lot of teams want memory that lives on their own disk, costs nothing per call, and never leaves the building. For agents handling anything sensitive, that second bet looks strong. Repo at github.com/TencentCloud/TencentDB-Agent-Memory.
Comments