TencentDB Agent Memory
An agent memory engine open-sourced by Tencent Cloud under MIT. It works on two fronts. Short term, full tool logs are offloaded to side files and all that stays in context is a task map written in Mermaid syntax; every node carries an id, and the agent drills back down to the raw text whenever it needs to check a detail. Long term, four tiers running from raw to synthetic: the conversation, the isolated fact (preference, constraint, event), the scenario filed per project, then the stable persona. Storage defaults to local, SQLite with its vector extension, with no external API call required.
Strengths
- Context keeps a readable task map instead of a pile of tool output
- Every node carries an id, so nothing is lost and the raw text stays reachable
- Runs locally by default on SQLite, with no mandatory dependency on a third-party service
- MIT licensed, so modifiable and integrable without negotiation
Limitations
- No gain on a short session: the machinery only pays off over duration
- One more local database to install, back up and migrate between versions
- The quantified gains come from the publisher's own benchmarks, to be rechecked on your own usage
Best for
- Long agent sessions where tool output saturates the context before the task ends
- Teams who want agent memory without shipping their data to a vendor
- Comparing a symbol-graph approach against fact-extraction memories