Auto-session -> respawn when tokens are exhausted.

[jmvl]

When Claude Code sessions approach the ~200K token context limit:

1. Performance degrades as context compression increases
2. Claude may lose track of earlier conversation context
3. User must manually recognize the problem
4. User must end session and start new one
5. New session loses working context (todos, in-progress work)
6. User must manually ask claude-mem to retrieve the last session context (and specify which one is the one!)

Impact

Issue	Severity	Frequency
Lost work context	High	Every long session
Manual intervention required	Medium	Every long session
Cognitive overhead	Medium	Constant during long sessions
Incomplete task handoff	High	~30% of respawns

Root Cause

Claude Code does not expose token usage metrics, and claude-mem has no mechanism to:

1. Detect approaching limits
2. Capture session state for handoff
3. Spawn continuation sessions
4. Inject handoff context

---

Core Architecture

1. Token Monitor Service - Estimates token usage via tool execution heuristics (since
   exact counts aren't exposed)
2. Respawn State Table - SQLite table storing TodoWrite items, last prompt, in-flight
   work
3. Enhanced Context Hook - Detects respawn state and injects continuation context
4. Respawn Orchestrator - Coordinates graceful handoff and new session spawn

Key Features

• Threshold alerts at 85%, 90%, 95% capacity
• Graceful handoff - Lets Claude complete atomic operations
• TodoWrite continuity - Preserves task list across sessions
• Auto-spawn - Starts new session automatically (configurable)

Integration Points

Uses existing hooks infrastructure (context-hook, summary-hook, save-hook) with
minimal new code.

claude-mem-auto-respawn-spec.md

[thedotmack]

Hi there! Thank you so much for this, I really appreciate the input and thought behind context management.

I read through this, and did some poking around, as it turns out there are some things in the above doc that aren't aligned with how Claude Code actually works. (side note did you see the setting and docs about my "endless mode" experiments?)

But that's not really a big deal, because Claude-Mem basically does what you want already, and i'm in mid-update to give more control to users for session start context options.

Also I use this tool for status line that tells me available context:
https://github.com/sirmalloc/ccstatusline/tree/main

I'm adding options so you can expand to full observations in session start and that would be like the "endless" continue, but token counts will be higher, so you'll probably want to have less full observations.

I also have a thought that I might want to make summaries also include the "semantic chunks" that I break out with observations, to see if we can pinpoint the best data to auto-include.

[thedotmack]

Another insight I just had -- I asked it to start saving the last message from the previous chat... but now that I'm typing this out, I dont think we even need that (I can grab it from the most recent transcript file)

you're watching my internal thinking rn 😂

[thedotmack]

this is happening now

[thedotmack]

I pushed an update that allows you to fine-tune customize your startup message.

I am including all kinds of goodies for you to play with @jmvl

Customize your context to your heart's content

Settings are in the UI at localhost:37777 in the sidebar

live preview is coming soon