Built three tools while building our AI assistant (OpenClaw) — kept hitting the same walls with agent context.

Relay: splits context into ephemeral (expires, re-fetch on startup) vs durable (stable, load directly). 30–50% fewer tokens, context is always current.

Stitch: async agent messaging with no broker. One agent drops a note, another picks it up on next startup. Flat files, zero infra.

PAG: entity graph with attention weights. Agents stop starting cold.

Full writeup with comparisons to mem0, Zep, LangMem, vector DBs: https://sparkco.ai/blog/agent-context-management-ephemeral-v...