06 — Chat Works. Does the Agent Actually Retrieve Memory?

The article explores whether an agent can actually retrieve memory, rather than just present information in context during a conversation. Initial tests showed that despite memory content appearing in the conversation context, the agent's tool execution environment might not be able to reach the memory addresses provided.

Two types of pathways exist for memory retrieval - passive injection, where the server side reads memory and includes it in context, and active retrieval using endpoints supplied in tool prompts. The article found that while passive injection continued working, active retrieval failed, even though the server's read path was functional.

Key factors were a missing configuration field called injection.externalGatewayUrl, which defaults to selecting an internal IP address when none is provided. This resulted in retrieval requests timing out or receiving no response, as the client tool execution environment was on an external network separate from the internal one where memory was stored.

The article stresses that merely seeing memory content in context does not prove retrieval functionality. Rather, it needs testing if the client can actually reach the memory address through proper configuration. Other probing results like HTTP authentication errors (401) or no response (000) also needed separate investigation, as these indicated issues in network connectivity and routing rather than memory storage itself.

In summary, while passive injection provided the conversation context with memory content, active retrieval proved problematic due to network configuration issues between the model's tool executor and the actual memory storage location. Fixing the missing configuration parameter resolved the active retrieval failure, but other independent issues related to knowledge resource URLs still needed separate verification.