spike(memory): evaluate local embeddings and vector search within Mate resource budgets

[robinbraemer]

Outcome

Evaluate whether fully local embeddings and vector similarity materially improve AgentOS private-memory recall inside persistent Mate resource budgets.

Scope

• Compare small local embedding models, including the Qwen3 embedding family, and compatible runtimes such as llama.cpp bindings or a narrowly scoped local process.
• Compare in-process cosine search, sqlite-vec, and no-vector baselines.
• Measure model/image size, architecture support, cold start, indexing time, query latency, memory/CPU pressure, index size, and retrieval quality.
• Define model/version change rebuilds and safe degradation when native extensions or model files are unavailable.
• Use synthetic or explicitly approved evaluation content only; do not upload real memory to a remote provider.

Acceptance criteria

• The decision record selects or rejects local vector retrieval with evidence.
• Candidate binaries/models have pinned provenance, licenses, checksums, and supported architectures.
• sqlite-vec/native-extension failure degrades safely and is visible.
• No vector result can bypass topic validation, privacy, scope, attachment, pause, or forget checks.
• A vector path is not adopted unless it beats lexical/current baselines within declared Pod budgets.