1. Compacting, cleaning, and truncating terminal outputs in the somatosensory transducer (thanks TokenJuice for the idea)
2. Working memory buffer is capped at 45k characters, but when it starts to exceed, it will chunk the stalest memories, compress then, them reinject them to the context. No amnesia needed.
3. Context is pinned to top of requests for large tool calls.
4. Passing between agents / tools is optimized to give directions for reading files, as well providing a topology search, so whole files don't have to be injected into context.
5. The memory system engages during sleep and dreaming, rotating logs and extracting anything important from them into memory specific domains (md files per domain).
I'll be continually improving this and would love ideas / criticism!
1. Compacting, cleaning, and truncating terminal outputs in the somatosensory transducer (thanks TokenJuice for the idea) 2. Working memory buffer is capped at 45k characters, but when it starts to exceed, it will chunk the stalest memories, compress then, them reinject them to the context. No amnesia needed. 3. Context is pinned to top of requests for large tool calls. 4. Passing between agents / tools is optimized to give directions for reading files, as well providing a topology search, so whole files don't have to be injected into context. 5. The memory system engages during sleep and dreaming, rotating logs and extracting anything important from them into memory specific domains (md files per domain).
I'll be continually improving this and would love ideas / criticism!