Claude Code

1. When a session starts

   • SessionStartSets the namespace and loads the skill

     • Resolves the namespace from ~/.memorylayer/namespaces.json and exports MEMORYLAYER_NAMESPACE, so every Bash command sees it
     • Names the session after the project folder and watches package.json, go.mod, Cargo.toml and friends
     • Reloads the /memorylayer skill and reports whether the code index is up to date

     ~/.claude/hooks/memorylayer_session_start.js
   • CwdChangedFollows you into another project

     • Re-resolves the namespace for the new directory and updates MEMORYLAYER_NAMESPACE

     ~/.claude/hooks/memorylayer_cwd_changed.js

2. Every prompt

   • UserPromptSubmitPuts memory in front of Claude before it answers

     • First turn: loads your rule, constraint and guide memories from every namespace, as content rather than a reminder
     • Every turn: tells Claude to recall from the project namespace and which tool fits the question
     • Ingest guard: skips re-indexing when the git commit hasn't changed

     ~/.claude/hooks/memorylayer_user_prompt_submit.js

3. Around every tool call

   • PreToolUseMemory first, then filescan block

     • The first Read, Grep, Glob or Bash of a turn is denied until recall() or code() has run
     • The retry is always allowed, so a turn can't get stuck when memory has nothing
     • Before Edit or Write it asks for code(impact) and a check for known errors

     ~/.claude/hooks/memorylayer_pre_tool_use.js
   • PostToolUseLifts the gate and nudges good habits

     • After recall or code: marks memory as consulted, which opens the PreToolUse gate for the turn
     • After edits: reminds Claude to track errors as one memory each, open → fixed

     ~/.claude/hooks/memorylayer_post_tool_use.js · memorylayer_mark_used.js
   • PostToolUseFailureTurns a failed tool call into an error memory

     • Checks whether the error is already known, opens [error][open] if not, and closes it as [error][fixed] once solved

     ~/.claude/hooks/memorylayer_post_tool_use_failure.js

4. When Claude finishes a reply

   • StopCheck the claims, save the learnings

     • Reminds Claude to run verify() on facts it just stated and to save anything new
     • Fires at most once per turn, so short replies don't get nagged
     • StopFailure fires every time — each failure is worth an error memory

     ~/.claude/hooks/memorylayer_stop.js

5. When the context fills up

   • PreCompact · PostCompactSave before the trim, reload after

     • Before: asks Claude to write down decisions and open errors — nothing else survives the trim
     • After: loads priority memories straight back in, because the reminder to recall was trimmed too

     ~/.claude/hooks/memorylayer_compact.js

6. Subagents and tasks

   • SubagentStart · SubagentStopSeed the helper, keep what it found

     • Start: gives the subagent the memories that match its task, since it begins with none of the conversation
     • Stop: asks Claude to keep any fix, pattern or decision the subagent surfaced

     ~/.claude/hooks/memorylayer_subagent.js
   • TaskCreated · TaskCompletedRecall past attempts, save the outcome

     • Seeds the task with matching memories, then saves the result and closes related error memories

     ~/.claude/hooks/memorylayer_task.js
   • TeammateIdleCheap upkeep while an agent waits

     • Suggests consolidate({ action: "ripple" }) or an explore pass — optional, skipped when there's nothing open

     ~/.claude/hooks/memorylayer_teammate_idle.js

7. When the session ends

   • SessionEndThe closing checklist

     • Runs in the background so exit isn't delayed: save guides, close errors, record the index, run consolidate()

     ~/.claude/hooks/memorylayer_session_end.js

Every hook reads or writes one store~/.memorylayer/memories.db