When Experience Becomes a Skill

TencentDB Agent Memory uses a different standard for a Skill than for an ordinary memory. The question is not merely whether a conversation contains a useful fact, but whether it contains a repeatable way of working that another Agent can execute under the right conditions. This chapter follows that conversion from conversation buffer to reviewed, versioned, and governed Skill, so the reader can decide when a procedure deserves its own asset and why the system does not turn every transcript into instructions.

A structured memory says something about what happened, what a user prefers, or what a team decided. A Skill says how to perform a class of work. That difference raises the cost of being wrong. A stale preference may make an answer less personal; a stale release procedure can cause a failed deployment. TencentDB Agent Memory therefore gives Skills their own extraction trigger, resource lifecycle, ownership checks, version checks, and loadout path.

A conversation must earn archival

The Skill path begins with a normalized conversation slice, not with the entire unfiltered request stream. The Proxy collects the relevant user, assistant, and tool messages for a session and sends them to Core’s conversation-add path. The handler maintains a buffer and metadata for that session, including accumulated bytes and the number of tool calls. It counts an initiated tool_call, not its paired tool_result, because counting both would make one Agent action look like two.

The buffer is not archived on every turn. It is archived when the accumulated work crosses a configured condition: enough tool calls, enough bytes, a request that needs compression, or an oversize case that requires a fallback strategy. A force-archive operation can also flush the current session buffer. These conditions are signals that the conversation may contain substantial work, not proof that a valid Skill is present. The later review step can still decide that the material is not reusable or that an existing Skill should be updated instead of creating a duplicate.

The size path explains why compression belongs to the Skill lifecycle. Tool-heavy coding conversations can become too large for a single review prompt, while a short sequence of well-defined tool calls may be sufficient. The handler prepares a bounded archive payload, preserving message roles and the identifiers that connect tool calls to results. If the ordinary buffer grows beyond its limit, the oversize strategy creates a usable review input rather than silently dropping the session. The system is managing an input to a later judgment, not claiming that compression preserves every nuance of the original exchange.

Ordering is the first important reliability decision. Core writes the archive before it registers the worker task that will review it. The reason is a race between a producer and a worker: if the task becomes visible first, a worker can look for the archive, fail to find it during a slow object-storage write, and discard a real task as a ghost.