Show and tell: four delivery waves with shared memory and much less correction churn

[fuzzymoomoo]

I''ve been using MemPalace in a real software-delivery workflow experiment, not just as a note store.

I ran a four-wave implementation slice on a project called Hades, with Codex and Claude working in series rather than in parallel. The higher-context planning and review work was done by a stronger model, and the bounded implementation slices were then handed to a lower-cost model.

The continuity primitive was simple:

• durable artifacts saved in repo
• a handoff capsule naming current state, source files, decisions, and the next bounded task
• a MemPalace refresh so the latest state became retrievable
• a cold-start wake-up pass for the next actor

What changed was that the incoming actor could start from shared memory plus authoritative artifacts instead of replaying the full chat transcript.

The strongest safe result was this:

• the run reached Wave 4 before the first manual smoke pass without the usual between-wave corrective debugging loop
• later testing still found issues, but they were bounded and understandable, which felt like a much better failure mode

The MemPalace-specific value here was:

• continuity across session boundaries
• less repeated rediscovery
• better cold-start resumption
• more viable mixed-cost multi-agent workflows

This work sits inside a broader CDD workflow I am building, but the relevant point here is simpler: MemPalace behaved more like continuity infrastructure than just memory storage.

Public-safe write-up:

• https://github.com/fuzzymoomoo/cdd-mempalace/blob/main/examples/hades-four-wave-delivery-experiment.md

Related handoff case study:

• https://github.com/fuzzymoomoo/cdd-mempalace/blob/main/examples/hades-shared-memory-handoff.md

Curious whether others are using MemPalace this way for handoffs, resumable work, or multi-agent continuity.