claude-code - 💡(How to fix) Fix [PROPOSAL] Expose compact/session lifecycle hooks for external memory layers [3 comments, 2 participants]

Problem

Five open issues request persistent memory (#14227, #32627, #34192, #34556, #46138). The community is already building solutions — 3-tier markdown architectures, knowledge graphs, structured memory layers. Each re-invents transcript access and compact interception because Claude Code doesn't expose the lifecycle events external memory needs.

The platform doesn't need to build persistent memory. It needs to expose the seams so the community can build it composably.

Proposed hooks

1. PreCompact — fires before context compaction

// settings.json
{
  "hooks": {
    "PreCompact": [{
      "type": "command",
      "command": "bella save --auto"
    }]
  }
}

What it enables: External memory layers save structured state before compaction flattens the context. Currently, if compaction happens between user messages, unfiled knowledge is lost (as documented in #34556 — 59 compactions, each a potential knowledge-loss event).

What the hook receives: Current session ID, token count, transcript path.

What the hook can return: Optional additionalContext to inject into the post-compaction context (e.g., a structured summary from the memory layer, replacing the lossy default compaction).

2. PostCompact — fires after compaction completes

What it enables: Memory layers can inject recalled context into the fresh post-compaction window. Different from PreCompact — this is the retrieval side, not the save side.

What the hook receives: Session ID, new token budget available, compaction summary.

3. SessionEnd — fires when a session terminates

What it enables: Final save opportunity. Currently, if the user closes the terminal or the process exits, no save happens.

What the hook receives: Session ID, transcript path, exit reason.

4. SessionStart — fires on session initialization (before CLAUDE.md)

What it enables: Memory layers inject recalled context before the first user message. Currently achievable via UserPromptSubmit hooks but semantically wrong — session start is the right lifecycle event.

What the hook receives: Session ID, project root, cwd.

Prior art

• Claude Code's existing UserPromptSubmit hook is used by multiple memory projects as a workaround for SessionStart, but it fires on every message, not just session init.
• The PreCompact hook is specifically requested in #31845.
• Bella is an open-source belief hypergraph memory layer that would immediately use all four hooks. Currently works around the missing lifecycle by intercepting transcripts post-hoc.
• The theoretical basis is Recursive Emergence — a framework for how beliefs should emerge, reinforce, and decay in bounded-memory systems.

What this does NOT propose

• No opinion on what memory should look like (flat files, knowledge graphs, hypergraphs — that's the community's job)
• No changes to /compact behavior itself
• No new built-in memory system

This is purely about exposing lifecycle seams so the five existing community efforts can compose with Claude Code instead of fighting it.