Overhauling the memory layer

Viewing the core of the problem not as the model but as the memory layer, the memory structure was completely redesigned.

• First, raw capture was set up to accumulate all context as-is in daily logs, reducing loss of conversational context.
• Once a day, a cron job compresses the logs and moves them into long-term memory, keeping only decisions, preferences, and stable facts while discarding noise in the process.
• Instead of storing memory as large chunks, it was broken down into an atomic structure following the 1 concept = 1 file principle. This approach reduced retrieval problems the most.
• Each file includes a lightweight graph layer that explicitly links related concepts, creating memory exploration paths using nothing more than markdown links.
• To reduce the problem of retrieval frequently missing things, each file includes synonyms, multiple phrasings, and keyword expansion to compensate for semantic mismatch.
• Finally, a self improvement loop was attached to separately log retrieval failures or omissions, allowing memory gaps to be corrected over time.

In the end, the key wasn't more memory, but a memory system structured to be reusable.