Checkpoint: What Actually Reaches the Next Turn?
At this checkpoint, TencentDB Agent Memory is best tested with a failure rather than a feature tour. Suppose a user asks how a previous release problem was resolved, but the current Agent gives a generic answer. The practical question is not simply whether “memory works.” It is whether the original turn was captured, whether its derived views are ready, whether retrieval selected it, whether the session was initialized in the right scope, and whether the context arrived as injected text or an available tool. Connecting those boundaries gives the reader a diagnostic model for the system already established.
The model below does not add another memory layer. It connects the timing of capture and consolidation with the ranking of evidence and the transport of context into an Agent request. That connection supports a useful judgment: a missing answer can have several causes that look identical at the user interface.
Read the timeline, not just the result
Start with the moment the earlier work occurred. A normal turn is captured into L0 after the Agent responds. The capture path is incremental and sanitized, so the stored evidence is not necessarily a byte-for-byte copy of the entire prompt. Harness wrappers and injected instructions can be removed, while the user’s real request and the assistant’s useful response are retained. If the session cursor or identity is wrong, the evidence may be missing, duplicated, or assigned to the wrong scope before any search begins.
If L0 exists, that still does not mean the structured memory is ready. L1 extraction is an asynchronous step. It applies a stricter quality gate, asks an LLM to identify durable records, and may legitimately produce no memory from a short or transient exchange. A successful capture and an empty extraction are different states. So are a valid L1 candidate waiting in a queue and an L1 call that failed to parse its output. Looking only at the current prompt cannot distinguish them.
The next transition is from L1 to the scene and profile views. L2 organizes related records into a coherent scene, and L3 creates a compact higher-level orientation. These views are useful precisely because they compress detail, but their usefulness arrives later than L0 capture. Timers, queues, and checkpoint state govern that delay. A scene can therefore be stale while its source records are current, or a profile can be absent while the underlying scene exists. “The Agent did not see the profile” does not establish that “the system did not record the conversation.”
Recall adds another branch. The current question is searched against the available L1 records using keyword, vector, or hybrid signals. A record can exist and be ready but fail to rank high enough, fall outside the isolation filter, or lose to a result cap. Conversely, a record can rank highly because its wording is similar while still being stale or wrong. Retrieval position is evidence for selection, not proof of correctness.
Finally, the result must enter a live client request. L3 and the L2 scene index may be placed in stable system context, while current L1 matches are put beside the user’s turn and deeper material remains behind a tool. If the session did not initialize, the Proxy may bypass injection.