Decide

Decide retrieves evidence, surfaces open conflict pairs, and asks a judge to recommend a resolution. The judge is the local Laya System-1 sidecar when [laya] enabled (one forward pass, ~100–200 ms on MLX for a batch), falling back to your coding-agent CLI. Hybrid search stays local; the Laya path needs no LLM at all.

When to use it

• Two decisions disagree (e.g. “SQLite forever” vs “move to Postgres”)
• An agent should pick a side with cited observation ids
• You want an optional recorded follow-up decision note

Usage

statefulmemory decide "Should we keep SQLite or move to Postgres?"
statefulmemory decide "Which auth strategy is in force?" --limit 12

MCP equivalent: memory_decide.

Pin the model if you are using the agent-CLI fallback:

export STATEFULMEMORY_LLM_PROVIDER=opencode
export STATEFULMEMORY_LLM_MODEL=qwen
# or STATEFULMEMORY_LLM_BIN=/path/to/cursor-agent

With Laya enabled, statefulmemory decide takes the fast path and no STATEFULMEMORY_LLM_* is needed. Raise STATEFULMEMORY_LAYA_TIMEOUT_MS=250 so decide batches (~100–200 ms) fit. See Laya System-1.

What you get

• A recommendation + short rationale
• Evidence observation ids
• Conflict pair status (open / resolved) when relations exist

Exact JSON/text fields depend on --output; use JSON in pipelines.

How it fits

Decide reads the same store as search & context. Conflict edges come from the supersession / relation graph built when conflict.enabled is on (Config). Record lasting outcomes with obs save.

Troubleshooting

Symptom	Fix
Fails immediately	No Laya and no agent CLI → install OpenCode/Cursor/Claude/Gemini or set STATEFULMEMORY_LLM_BIN
Thin evidence	Save more notes; raise --limit; try a clearer question
Want local-only	Enable Laya via smem install, or disable conflict.enabled / extract.enabled and keep hybrid search