Memory gauge forces chat termination at 0% while 60%+ token budget remains unused - rendering majority of paid tokens unusable

Issue Summary

Claude Code's memory gauge reaches 0% and forces chat termination when 60-65% of the token budget remains available, making the majority of purchased tokens unusable and severely disrupting workflows.

Severity: High

Impact:

• 60% of token budget becomes completely unusable - paying for 200k tokens but can only use ~80k
• Forced workflow interruptions despite ample resources remaining
• Unnecessary context loss requiring premature handovers
• Productivity impact - complex multi-stage tasks interrupted mid-work

Current Behavior (Actual)

1. At ~35% token usage (~70k/200k tokens used):

• Memory gauge shows "10% remaining"
• Warning appears: "Memory at 10% - consider using /compact"
• Token budget correctly shows 65% remaining (130k tokens available)

1. At ~40% token usage (~80k/200k tokens used):

• Memory gauge reaches 0%
• Chat freezes completely
• Token budget still shows 60%+ remaining (120k+ tokens unused)
• Only options: /compact (loses context) or start new chat

1. Result:

• Cannot use remaining 60% of token budget
• Forced to abandon chat and start new session
• Context loss disrupts complex workflows
• Wasting 120k+ tokens per session

Expected Behavior

1. Memory gauge should align with token budget

• If 60% tokens remain, memory should reflect similar capacity
• Both gauges should reach 0% at approximately the same time

1. Chat should remain functional until token budget depleted

• Should be able to use full 200k token allocation
• No premature freezing with resources available

1. Warnings should be accurate

• Memory warnings should match actual resource availability
• Should only force termination when genuinely out of resources

Reproduction Steps

1. Start a new chat with Sonnet 4.5 (200k token budget)
2. Engage in complex, multi-stage work (e.g., Issue #50 workflow with multiple subagents)
3. Monitor both indicators:

• Memory gauge (bottom of interface)
• Token budget counter

1. Continue working until memory shows 10% warning
2. Note token budget still shows 65%+ remaining
3. Continue until memory reaches 0%
4. Observe: Chat freezes, but token budget shows 60%+ remaining

Frequency: Consistent - happens in every long session

Evidence from Recent Session

Session just completed (2025-11-04):

• Work: Issue #50 - Modular docs restructure (complex multi-agent workflow)
• Token usage at handover: ~84k/200k (42%)
• Memory gauge status: "Consider using /compact" warnings appearing
• Result: Forced to create checkpoint and start new chat despite 58% tokens remaining

Chat log excerpt:

<system_warning>Token usage: 83914/200000; 116086 remaining</system_warning>

116k tokens remaining, yet memory warnings forcing chat termination

Technical Details

Environment:

• Model: Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
• Token budget: 200,000 tokens
• Platform: Windows (win32)
• Claude Code version: Latest as of 2025-11-04

Usage Pattern:

• Complex multi-stage workflows with Issue tracking
• Multiple subagent launches (Task tool usage)
• Structured development process with checkpoints
• Typical session: 80k-100k tokens used before forced termination

Business Impact

Resource Waste:

• Paying for 200k tokens but can only use ~40% before forced restart
• 60% of purchased capacity is unusable per session
• Equivalent to paying for a resource you cannot access

Workflow Disruption:

• Complex tasks (like Issue #50: 47 file creation, verification, cross-referencing) get interrupted mid-work
• Forced handovers create risk of context loss
• Additional overhead: creating checkpoints, handover documentation, new chat startup
• Reduces productivity - cannot complete work in single session

User Experience:

• Frustrating to see 120k+ tokens available but chat frozen
• Confusing mixed signals from UI (memory: 0%, tokens: 60%)
• Undermines confidence in resource indicators

Additional Context

What makes this worse:

• Cannot plan around it - memory gauge does not correlate to token usage predictably
• /compact may temporarily help but loses valuable context
• Starting new chat requires extensive handover documentation to maintain continuity
• Particularly impacts structured workflows with multiple stages

Suspected Root Cause:

• Memory gauge appears to track something other than token usage
• Possibly: conversation history length, subagent spawns, or context window separately from tokens
• Regardless: these should be aligned or clearly labeled as separate limits

Suggested Fix

1. Align memory gauge with token budget

• Both should reach depletion at same time
• Or clearly label if they track different resources

1. Increase memory capacity to match token budget

• If memory is separate limit, it should support full 200k token usage

1. Improve warning accuracy

• Only show "low memory" when actually running out of usable capacity
• Do not freeze chat while significant token budget remains

1. Better documentation

• If memory and tokens are intentionally separate, document this clearly
• Explain what memory tracks vs tokens

Workarounds Currently Using

1. Proactive handovers - stop at 40% token usage instead of continuing
2. Extensive checkpoint documentation - prepare for forced restarts
3. Structured Issue tracking - document everything in GitHub Issues for continuity
4. Avoiding long sessions - break work into smaller chunks despite overhead

These workarounds reduce productivity and waste paid resources.

---

Request

Please investigate the misalignment between memory gauge and token budget. Users should be able to utilize their full token allocation without premature chat termination.

This issue makes 60% of each chat session unusable despite resources remaining.

Thank you for looking into this.