Deep Research Ldr Automem

In other words:

• the research layer decides whether to recall
• AutoMem decides what memory items are most relevant

Operating rules

1) Start with the research question

Write down:

• the primary question
• sub-questions
• freshness requirement
• whether internal history matters

2) Inspect repo-local evidence first

Before widening scope:

• inspect repository structure
• inspect current code / docs / tests
• inspect local design notes if present

3) Use conditional memory gating

Do not call memory by default for every task.

Prefer memory when the task contains signals like:

• previous / earlier / last time
• decision / tradeoff / rationale / why
• preference / pattern / habit / style
• internal / repo / team / project
• follow-up research on an earlier thread

Prefer skipping memory when the task is mostly:

• public factual lookup
• latest / current / status questions
• fresh external news or release tracking

4) Suggested decision modes

• forced — explicit memory lookup requested
• strong — memory is clearly valuable
• light — memory may help, keep it narrow
• skip — no memory lookup

A practical threshold policy is:

• forced when the contract explicitly says so
• strong when memory relevance is high
• light when relevance is moderate
• skip otherwise

5) AutoMem read policy

When memory is used:

• use the AutoMem API recall surface
• do not query memory backends directly
• preserve memory score / provenance
• keep url and link when the downstream system distinguishes them

For stronger context:

• sort=score
• expand_relations=true
• reasonable expansion filters
• explicit context_types

For lighter context:

• smaller limits
• expand_relations=false

Do not start with aggressive entity expansion.

6) Blending policy

Treat memory as a memory plane, not as a replacement for public research.

Use:

• memory for internal continuity
• public sources for fresh / external verification
• synthesis to compare them

7) Freshness rule

If the task is freshness-sensitive:

• public search is mandatory
• memory can shape context or preferences
• memory is not the source of truth for current public state

8) Failure rule

If memory lookup fails:

• record the failure
• continue with repo inspection + public research
• do not fail the whole research run by default

Deliverables

At minimum, leave behind:

• .agents/research-plan.md
• .agents/research-findings.md
• source ledger
• final recommendation
• explicit residual uncertainty
• follow-up implementation brief or issue drafts when relevant

Final output discipline

Separate these clearly:

• confirmed fact
• repository observation
• memory-derived context
• public-source evidence
• inference
• recommendation

Do not overstate certainty. Do not present memory-derived content as fresh public truth. Do not silently cross from research into code changes without checking the contract.