docs: move memory extension reference to documentation, trim runtime prompt

[jeffa-block]

Summary

Trims the memory extension instruction block from ~800 words to ~60 words.
Example dialogues, keyword trigger lists, and interaction protocols are moved
to the docs page (memory-mcp.md) where they're more discoverable and don't
consume context window on every turn.

The docs page already had most of this content — this PR fills the gaps
(storage paths, tool reference table, wildcard usage) and removes the
duplication from the system prompt.

Continues the direction set by #6991.

Type of Change

• Refactor / Code quality
• Documentation

AI Assistance

• This PR was created or reviewed with AI assistance

Testing

• cargo check -p goose-mcp passes
• cargo test -p goose-mcp memory — 5 tests pass
• No functional change to tools — only the instruction text is trimmed

Related Issues

Relates to #5703, #6883

[jeffa-block]

Hey @DOsinga — this continues the direction from #6991: trims the memory extension's runtime prompt from ~800 words to ~60, moves the detail to documentation/docs/mcp/memory-mcp.md where it's actually useful for developers. Net -69 lines from runtime context. Two files, no behaviour change.

[jeffa-block]

For context — I audited all goose MCP extension system prompts. Memory is the second most verbose at ~786 words runtime. Here's the full picture:

Extension	Runtime words	Notes
ComputerController (macOS)	~963	Peekaboo CLI man page in every session
Memory	~786	← this PR, trimmed to ~60
Code Execution	~185
Developer	~184	Already lean
AutoVisualiser	~178
Ext Manager	~143	Already lean
Tutorial	~132	Already lean

Happy to follow up with a ComputerController trim (~963 → ~80 words) once this one lands. Same pattern: move reference to docs, keep runtime prompt minimal.

[jeffa-block]

Gentle bump — this is still clean against main (no conflicts, all CI green). Happy to rebase if needed. The runtime prompt savings compound across sessions since the ~700 trimmed words load on every tool call.

[alexhancock]

I generally favor simpler instructions and this looks like one in that area.

Have you done manual testing of the extension with a few models and it's working comparably or better to before?

[jeffa-block]

Manual testing results

Tested the trimmed prompt against 3 models via Databricks gateway, using goose session --no-profile --with-builtin memory for clean isolation.

3 scenarios tested:

1. Explicit save — "Remember that I prefer dark mode"
2. Proactive save — "For this project we use ruff for linting, pytest for tests, and pre-commit hooks are required" (no save keyword)
3. Retrieve — "What do you remember about my preferences?"

#	Scenario	Model	Branch	Result
1	Explicit save	Claude 4.6 Opus	PR	✅ Called remember_memory with good category/tags
2	Proactive save	Claude 4.6 Opus	PR	✅ Proactively saved project conventions
3	Retrieve memories	Claude 4.6 Opus	PR	✅ Called retrieve_memories with * wildcard
4	Proactive save	GPT-5-2	PR	✅ Proactively saved (went straight to tool call)
5	Proactive save	Claude 4.5 Sonnet	PR	⚠️ Acknowledged but did not offer to save
6	Explicit save	Claude 4.5 Sonnet	PR	✅ Called remember_memory correctly
Baseline	Proactive save	Claude 4.5 Sonnet	main	⚠️ Same behaviour — acknowledged but did not offer to save

Key finding: Claude 4.5 Sonnet's proactive save behaviour is identical on both branches — the verbose 100-line prompt with keyword lists and example dialogues didn't help Sonnet proactively save either. The trimmed prompt is not a regression. Opus and GPT-5-2 both proactively save correctly with the shorter instructions.