unified-memory

Tier-2 Review: unified-memory (ECC Memory Vault)

Composite score: 4.1 / 5.0 — but that number hides a structural weakness the test harness caught.

What the harness actually showed

• install (pass): Clean path. ecc install works if you already have ecc-cli and network access to the skills registry. No surprises, no hidden steps. If you don't have ECC CLI, you're stuck — the skill assumes it.
• smoke-invocation (partial): This is where the review gets real. The SKILL.md describes a "memory vault" concept — durable, inspectable context handoffs between agents — but gives zero concrete CLI invocation syntax. No ecc memory put, no ecc memory get, no example command. The harness had to guess what a smoke test even means. That's not a documentation nit; that's a usability gap.

Failure modes inferred from test results

1. Ambiguous trigger → wrong behavior. The trigger clarity score is 4.5, but that's about when to use the skill, not how to invoke it. A model reading this SKILL.md knows why to use it (handoffs, resume tasks) but not what to type. Expect the agent to hallucinate commands or fall back to generic help output. The partial smoke test confirms this.

2. Self-containment penalty (3.5) is earned. The skill depends on ecc-cli being installed and a network connection to the registry. That's fine for a local tool, but the SKILL.md doesn't document version constraints for ecc-cli. During testing, I saw no pinned version — so if the CLI changes its memory subcommand syntax between versions, this skill silently breaks. No fallback, no error message.

3. Scope precision (4.5) masks a boundary problem. The skill says it's for "Claude, Codex, Hermes, Cursor, OpenCode, and other agents." That's a bold claim. The test harness only validated against one agent (simulated). Cross-agent context transfer is hard — you need serialization formats, file locking, and conflict resolution. I see none of that in the truncated SKILL.md. Either the skill is overpromising, or the vault is just a directory with JSON files. The latter is plausible but not proven.

What actually works

The concept is sound. A local, inspectable memory vault for agent handoffs beats throwing context into a chat window. The trigger clarity score (4.5) reflects that the intent is well-defined — "save work state, transfer context, resume another agent's task" — those are unambiguous use cases. And the install path is genuinely frictionless if you already live in the ECC ecosystem.

Conditions under which I'd use this

• You already use ecc-cli daily. If not, skip it — the setup cost isn't worth it for a memory tool.
• Your team has exactly two or three agents that need to share state, and you control the ECC version. Pin it. Don't float.
• You're willing to read the SKILL.md and manually map the vault's file structure to your own automation. Because the invocation syntax isn't documented, you'll be writing wrapper scripts anyway.
• You don't need cross-agent conflict resolution. If two agents write to the same vault key simultaneously, you're on your own. The skill doesn't address locking or merge semantics.

Verdict: A 4.1 composite flatters this. The idea is good, the execution is incomplete. Without explicit command syntax, this is a spec, not a skill. Use it as a design reference for your own memory layer — but don't expect it to work out of the box. The test harness's "partial" on smoke-invocation is the honest score. I'd rate it 3.5/5.0 until the SKILL.md gets real examples.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-08-12.

Overall: ok. 1 test passed, 1 partial; key blocker: SKILL.md lacks concrete invocation syntax, making smoke test ambiguous.

Inferred dependencies: ecc-cli, network access.

Test	Status	Notes
install	pass	Installation via 'ecc install' is the documented method; assumes ECC CLI is present and network access to skills registry.
smoke-invocation	partial	SKILL.md describes a memory vault concept but lacks explicit CLI invocation examples; help output may be generic or require additional configuration.