Deep Research Intake Notes (2026-05-20)

Source inputs:

• Copilot shared task response: https://copilot.microsoft.com/shares/tasks/ajvthvsd2irkDvch8TAQV
• Raw captured page: temp-deep-research-response

Purpose:

• Preserve the main external findings in a decision-ready format.
• Separate generally strong patterns from weaker or blog-only evidence.
• Feed the next MemorySmith council review with concrete questions and acceptance gates.

What This Resolves

The newly captured research largely supports the current RFC direction:

• Convention-first remains reasonable for local-first systems.
• UI validation/autocomplete is usually needed before metadata conventions scale.
• Staleness should be surfaced before applying hard suppression/decay.
• Hybrid output shape is preferred: JSON for tool interfaces, Markdown for human docs.
• Relationship typing should be promoted incrementally, not front-loaded.

This does not by itself decide MemorySmith-local defaults (roles, exact thresholds, compatibility choices). Those stay local decisions.

High-Value Findings (Most Actionable)

1. Tag governance should not remain free-form at scale.

• Practical pattern: keep conventions initially, then add UI assistance and validation before promoting to schema.
• Immediate implication for MemorySmith: prioritize tag chips/autocomplete/validation before ranking behavior depends on namespaced tags.

1. Staleness handling should be warning-first.

• Practical pattern: explicit freshness/review metadata and visible stale markers outperform silent drops.
• Immediate implication: add stale/superseded/expired warnings in search and context-pack output before any aggressive decay.

1. Page chunking should be evidence-triggered, not assumed.

• Practical pattern: start with whole pages for small corpora; chunk when page length and retrieval misses justify it.
• Immediate implication: define local trigger metrics (page length distribution, miss rate, recall@k regressions) before implementing chunking.

1. JSON default for agent/tool output is widely favored.

• Practical pattern: machine interfaces should return strict JSON; render Markdown summaries in UI.
• Immediate implication: preserve Markdown for pages/chat readability but bias MCP tool responses toward structured JSON for agent consumers.

1. Governance and approval need traceable evidence.

• Practical pattern: treat AI writes as proposals with provenance, confidence, and reviewable diffs.
• Immediate implication: for critical memory writes, require source-backed rationale and make approval rationale visible in trace/review records.

Evidence-Quality Caveat

The captured research blends source types:

• Stronger: platform docs, standards-style guidance, mainstream vendor docs.
• Moderate: engineering handbooks, architecture writeups, operational reports.
• Weaker: personal blogs, opinion-heavy posts, unreplicated claims.

Use this intake as directional evidence, not final proof. For each implementation decision, require at least one strong source or local benchmark.

Local Decision Questions Still Open

These remain MemorySmith-internal and should be settled by council review + local measurement:

• Exact tag canonical forms and which ones are mandatory.
• Exact stale behavior policy (warn vs demote vs filter).
• Exact page-chunking trigger threshold.
• Whether memorysmith_context_pack should switch defaults while preserving existing consumers.
• Exact RBAC requirements for Agent writes into Core memories or strict-rule content.

Recommended Next Council Packet

Use this packet when running the next council review:

• Core Memory System Improvements RFC
• AI Memory Suite Implementation Plan
• Council Workflow
• Deep Research Prompt
• temp-deep-research-response
• This intake note

Decision target for the next review:

• Use the implementation plan's Phase 0 maintenance audit as the first gate.
• Approve a Phase 1 implementation envelope limited to policy, diagnostics, validation, and warning surfaces.
• Defer ranking decay, schema expansion, and chunking until local metrics prove need.

Suggested Acceptance Gates (Pre-Implementation)

1. Tag validators and UI cues exist before namespaced tags affect ranking.
2. Stale/superseded warnings appear in search/context outputs and chat references.
3. A benchmark/probe suite exists for strict-rule retrieval and stale-context safety.
4. A compatibility note exists for any JSON-default output changes.
5. Critical Agent write approvals include explicit evidence/provenance in review surfaces.