Adapter example: connect Ollama or llama.cpp to the v0.24.1 shared memory protocol

[qh-work]

Current target

Use v0.24.1, not the retired v0.21 host envelope or unqualified main. The release now includes a generic visible-event adapter and configured protocol/MCP entry points. An Ollama or llama.cpp model server alone is not an agent harness and does not imply automatic lifecycle events.

Bounded contribution

Choose one already-authorized local agent runtime using either model server. Provide a small optional example that saves and recalls through the current protocol or generic adapter, without a model-specific database.

• Preserve canonical records, provenance and relations; sessions and tasks do not own memory.
• Show the host's actual input, final, cancellation and permission boundaries. Do not infer automatic capture from a manual command.
• Save visible user/final-assistant text only; no hidden reasoning, tool traces or transcripts.
• Use stable event IDs and exact retries; incomplete or cancelled turns must not become save confirmations.
• Distinguish local saving, optional signed sync, and remote AI consumption. Signatures identify keys, not truth or execution permission.
• Publish only synthetic examples and content-free results, with exact release, runtime version and skipped cases. Label unexecuted recipes honestly.

See the review handoff. No application tests were run by the maintainer for this release. Contributions require the contributor's current user/host authorization; no self-installation, log suppression or acquisition of new permissions is in scope.