Author here. I needed to build a memory layer for some enterprise agents and went looking for an existing open source project that did the things I needed. I noticed that the harness layer (Claude Code, Cursor, Copilot) is well-served by per-agent memory, and there are a lot of good projects for local-only or single-agent memory, but none that took care of memory provenance, contradictions, RBAC, filtering, etc., that were not part of a SaaS offering.
Enabling agent hive-mind and allowing for governance and controls seemed to be generally unaddressed as a platform concern.
Harrison Chase and Sarah Wooders have both recently argued that memory belongs in the harness. They're right for single-user, single-agent setups. This piece tries to articulate why memory becomes a platform concern at scale, and why I think this ends up as actual infrastructure (consistency, audit, retention) rather than another harness feature.
I’m genuinely curious what people running multi-agent systems in production are doing here. Treating it as solved with RAG? Bespoke per-team memory layers? Punting until something standardizes? Cloudflare's Project Think and the OpenViking release feel like early signals but neither is portable in the MCP-style "follows the protocol, not the platform" sense.
Author here. I needed to build a memory layer for some enterprise agents and went looking for an existing open source project that did the things I needed. I noticed that the harness layer (Claude Code, Cursor, Copilot) is well-served by per-agent memory, and there are a lot of good projects for local-only or single-agent memory, but none that took care of memory provenance, contradictions, RBAC, filtering, etc., that were not part of a SaaS offering.
Enabling agent hive-mind and allowing for governance and controls seemed to be generally unaddressed as a platform concern.
Harrison Chase and Sarah Wooders have both recently argued that memory belongs in the harness. They're right for single-user, single-agent setups. This piece tries to articulate why memory becomes a platform concern at scale, and why I think this ends up as actual infrastructure (consistency, audit, retention) rather than another harness feature.
I’m genuinely curious what people running multi-agent systems in production are doing here. Treating it as solved with RAG? Bespoke per-team memory layers? Punting until something standardizes? Cloudflare's Project Think and the OpenViking release feel like early signals but neither is portable in the MCP-style "follows the protocol, not the platform" sense.