What shared memory taught me about agent coordination

These lessons came from Velqua, a local-first memory experiment. The project page explains the public product boundary.

Running several independent coding and research agents exposed a basic problem: useful context remained trapped inside whichever session found it. Copying entire transcripts between agents was noisy, fragile, and difficult to correct.

Shared is not the same as trusted

Once multiple agents can read the same context, a mistaken conclusion can spread farther. Shared memory needs provenance and status so later agents can distinguish a verified result from a working note or a guess.

Relevance beats volume

Giving every agent every retained fact quickly makes the context less useful. Good coordination selects the smallest relevant packet for the current task and leaves the rest available for deliberate retrieval.

Ownership must remain clear

An agent can propose a note, but it should not silently turn that note into canonical truth. The source that owns the work should retain the ability to correct, supersede, or remove it.

Handoffs are better than invisible synchronization

The most dependable pattern was an explicit handoff: what changed, what evidence exists, what remains uncertain, and what the next agent should do. Shared storage helps, but the handoff contract is what makes the information usable.

The system I want is not one giant memory. It is a governed exchange where context has provenance, scope, an owner, and a way to be corrected.

This article intentionally withholds private ranking heuristics, agent identity rules, internal message formats, memory contents, and deployment details. If the governance problem overlaps with work you are doing, reach out.