Agent rules

The always-on instructions that make an agent check memory first. memorylayer setup writes them for you; copy them here for any other agent.

The rules, per agent

~/.claude/CLAUDE.md266 lines · 14 sections

You have persistent memory + code intelligence via MemoryLayer (MCP). The UserPromptSubmit hook injects instructions each turn — follow them. Skipping them means answering without project history, prior decisions, or code knowledge already stored. Use memory tools before reading files.

Every turn

Load schemas with ToolSearch (first turn only — deferred tools need this).
Run both recall calls in parallel (user + project namespace).
code(ingest) runs automatically on the first turn of each session only.
Before reading files, answer from memory — pick the right tool:
Specific lookup (symbol/function/fact): recall({ query, namespace, intent: "find_symbol" })
Broad/exploratory question: recall({ query, namespace, explore: true }) explore=true returns primary results PLUS related memories via tag, temporal, and semantic edges — use it when you want surrounding context.
Only fall back to Read/Bash if the above returns answer_kind="empty".
After significant work, save what you learned with remember.

Tool selection quick reference

recall: read anything: search · answer · weave · list · load (intent-routed)
remember: write anything: upsert · update · delete · chunk · batch (param-routed)
code: all code intelligence: ingest · locate · impact · grep (action-routed)
consolidate: maintenance + ACT-R compile: TTL prune · replay-and-derive (ripple) · compose chunk
verify: hallucination gate: split text into claims, check each vs stored knowledge
prove: formal proof for a single claim (proven > retrieved confidence)
domain: formal domain management: load · export · status · set_mode
audit: batch Z3 invariant checker — CI-ready

Which tool, when

WhenExact skill or symbol name known
recall({ item_type: 'skill' })O(1) deterministic — fastest path
WhenCode question about an indexed project
recall(answer)Primary entry — always first for code Qs
WhenSession just started
recall({ preview: true })Mandatory before first response
WhenBrief message / need recent context
recall(list)Recency-sorted, no search needed
WhenNew project mentioned
code(ingest)Mandatory, run once
WhenUser shares preference / decision / correction
rememberDedupes automatically
WhenBroad “how does X relate to Y” (Pro+)
recall({ explore: true })Linked memories, derived facts, open loops
WhenCode answer was empty — need raw hits
code(locate)Intent-based fallback
WhenNeed specific literal in code
code(grep)Fallback for literal verification
WhenLong document to store (Pro+)
remember(chunk)Splits into searchable chunks
WhenMultiple facts at once (Pro+)
remember(batch)Up to 100 items per call
WhenEnd of session — derive + prune
consolidate(ripple)Replays recent memories to derive transitive facts + surface contradictions

Something missing or wrong? Email support@memorylayer.in or open an issue.

The rules, per agent

~/.claude/CLAUDE.md266 lines · 14 sections

Every turn

Load schemas with ToolSearch (first turn only — deferred tools need this).
Run both recall calls in parallel (user + project namespace).
code(ingest) runs automatically on the first turn of each session only.
Before reading files, answer from memory — pick the right tool:
Specific lookup (symbol/function/fact): recall({ query, namespace, intent: "find_symbol" })
Broad/exploratory question: recall({ query, namespace, explore: true }) explore=true returns primary results PLUS related memories via tag, temporal, and semantic edges — use it when you want surrounding context.
Only fall back to Read/Bash if the above returns answer_kind="empty".
After significant work, save what you learned with remember.

Tool selection quick reference

recall: read anything: search · answer · weave · list · load (intent-routed)
remember: write anything: upsert · update · delete · chunk · batch (param-routed)
code: all code intelligence: ingest · locate · impact · grep (action-routed)
consolidate: maintenance + ACT-R compile: TTL prune · replay-and-derive (ripple) · compose chunk
verify: hallucination gate: split text into claims, check each vs stored knowledge
prove: formal proof for a single claim (proven > retrieved confidence)
domain: formal domain management: load · export · status · set_mode
audit: batch Z3 invariant checker — CI-ready

Which tool, when

WhenExact skill or symbol name known

recall({ item_type: 'skill' })O(1) deterministic — fastest path

WhenCode question about an indexed project

recall(answer)Primary entry — always first for code Qs

WhenSession just started

recall({ preview: true })Mandatory before first response

WhenBrief message / need recent context

recall(list)Recency-sorted, no search needed

WhenNew project mentioned

code(ingest)Mandatory, run once

WhenUser shares preference / decision / correction

rememberDedupes automatically

WhenBroad “how does X relate to Y” (Pro+)

recall({ explore: true })Linked memories, derived facts, open loops

WhenCode answer was empty — need raw hits

code(locate)Intent-based fallback

WhenNeed specific literal in code

code(grep)Fallback for literal verification

WhenLong document to store (Pro+)

remember(chunk)Splits into searchable chunks

WhenMultiple facts at once (Pro+)

remember(batch)Up to 100 items per call

WhenEnd of session — derive + prune

consolidate(ripple)Replays recent memories to derive transitive facts + surface contradictions