Context

PR #491 wired validate_agent_id into the three session-start surfaces so a malformed agent_id is loudly rejected:

mem_session_start(agent_id="foo:bar")
→ "Error: invalid agent-id 'foo:bar': must match [A-Za-z0-9._-]+ ..."

The same hostile shape, sent at the multi-agent registration / search surface, takes a different path:

mem_agent_register(agent_id="foo:bar")
→ silently runs sanitize_namespace_segment("foo:bar")
→ stores under sanitised ``agent-runtime:foo_bar`` (or similar)

(See packages/memtomem/src/memtomem/server/tools/multi_agent.py:71-75 and :118-123 plus packages/memtomem/src/memtomem/cli/agent_cmd.py:108-112.)

So the same input now produces two UX outcomes depending on which tool the caller picked first — registering an agent rewrites the id in place, but starting a session for that same agent rejects it. That asymmetry is small today (multi-agent surface review just shipped) but it'll get harder to reason about as more tools fan out.

What we want

A deliberate decision between two options, applied consistently across mem_agent_register, mem_agent_search, and the matching mm agent ... CLI subcommands:

1. Switch to validate. Replace sanitize_namespace_segment calls with validate_agent_id. Same loud-reject behavior as the session-start surfaces. Breaking for any caller that was relying on silent sanitisation, but those callers were storing under unexpected namespaces anyway.
2. Document the divergence. Keep the silent sanitise behavior, but add a docstring + user-guide note explaining that registration / search are forgiving and that mem_session_start is the strict gate.

Option 1 is the cleaner long-term answer; option 2 is the cheaper short-term answer. Either way, the choice should be visible in the docstring + CHANGELOG so users aren't surprised.

Acceptance

• One of the two paths above shipped, not "we'll figure it out later."
• If option 1 (validate): regression tests pin the error parity with the session-start surfaces; CHANGELOG flags the breaking shape.
• If option 2 (document): docstrings + docs/guides/multi-agent-quickstart.md (or equivalent) note the read/write asymmetry; the validator's docstring in constants.py stays honest about which surfaces it covers.

Out of scope

• The LangGraph adapter parity gap — tracked in its own follow-up issue.
• Adding a new "loose vs strict" config knob — that adds an axis we don't actually need until users ask for it.

Provenance

• PR fix: validate agent_id before agent-runtime namespace concat (CLI + MCP) #491 review (concern 2): mem_agent_register / mem_agent_search semantic divergence with the new session-start gate.
• Same review's verdict explicitly recommends filing this separately so it doesn't drift quietly.