Overview
TencentDB Agent Memory turns completed Agent work into governed, reusable context rather than a larger chat log. The series follows capture from sanitized L0 evidence through selective L1 memories, L2 scenes, and L3 orientation; retrieval then ranks and scopes what enters a protocol-correct session, while Skills, Wiki, and CodeGraph serve different kinds of reuse. The result can reduce repeated explanation and rework, but freshness, extraction accuracy, permissions, client behavior, and deployment state remain conditions—not guarantees.
Learning outcomes
- Layered memory trades detail for faster orientation: sanitized L0 evidence becomes structured L1 records, L2 scenes, and compressed L3 context.
- Retrieval combines keyword, embedding, or hybrid signals, but ranking indicates relevance rather than truth and remains bounded by identity, permissions, and prompt budget.
- Chat Memory, Skills, Wiki, and CodeGraph support different reuse jobs; choosing the asset by question avoids treating reference or procedure as universal instruction.
- MemoryProxy keeps client protocols recognizable while managing session identity, context injection, auxiliary-request bypass, and write-back; a successful response does not prove later processing completed.
- Standalone keeps data and pipeline state local; service mode shares storage, queues, and locks for multi-Agent coordination at the cost of operational dependencies.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
How is TencentDB Agent Memory different from storing a chat log?
It turns captured interaction into several reusable, scoped views instead of treating the transcript as one undifferentiated context. L0 preserves sanitized evidence, while L1–L3 provide increasingly selective records, scenes, and orientation.
What do the L0, L1, L2, and L3 memory layers do?
L0 preserves captured conversation evidence, L1 extracts structured records, L2 groups them into scenes, and L3 compresses scenes into profile or operating context. Each layer loses detail for a different retrieval job and remains a derived view except for the lower capture record.
How does TencentDB Agent Memory decide what an Agent receives?
It combines retrieval ranking with identity, asset bindings, visibility, ACLs, and context budgets before injecting material or exposing it through tools. Relevance alone cannot override scope, and deeper evidence may stay on demand.
When should an Agent use Chat Memory, a Skill, Wiki, or CodeGraph?
Use Chat Memory for past interaction, a Skill for repeatable action, Wiki for document-based rationale or guidance, and CodeGraph for code relationships and impact. These assets can cooperate, but each calls for a different verification path.
Should TencentDB Agent Memory run standalone or as a service?
Standalone fits local, single-process continuity; service mode fits shared assets or multiple replicas by externalizing storage, timers, queues, and locks. Service mode adds coordination, credentials, recovery, and monitoring work.