Context bloat 400k tokens in 3 min after restart

[varunbaranwal]

Issue Summary

Container restart does not clear accumulated session context. Running agent session continues to load massive context despite restart, resulting in 90%+ increase in API token costs.

---

Technical Details

Session Information:

• Session ID: sess-1778272236516-2bodft
• Agent Group: Prachin (Private Secretary Agent)
• Messaging Group: telegram-mg-17782 (Telegram)
• Session Started: May 8, 2026
• Container Status: Restarted May 19, 2026, 8:53 AM IST

Issue Timeline:

Date/Time	Event	Token Usage
May 8	Session created, initial usage baseline	~5-10K/msg
May 17	Context bloat warning identified	~100K/msg
May 19, 8:45 AM	Spike detected: 575K tokens in 3 min	~192K/msg
May 19, 8:53 AM	Container restarted (docker restart)	-
May 19, 8:58 AM	Post-restart spike: 397K tokens in 3 min	~132K/msg
May 20, 9:00 AM	Status: No improvement, continues at 132K/msg	132K/msg

---

Root Cause Analysis

Expected Behavior:

• Fresh session restart should clear context
• New messages should load minimal context (~5-10K tokens)
• Cost should drop 90%

Actual Behavior:

• Container restart completed successfully
• Session ID remains same: sess-1778272236516-2bodft
• Context size: Still ~130K tokens per message
• No cost reduction observed

Hypothesis:

1. Session context persisted in database/cache despite container restart
2. Agent loading accumulated context from storage layer, not fresh memory
3. Possible: inbound.db or outbound.db not cleared
4. Possible: Session container restarted but agent retained state

---

Business Impact

Cost Analysis:

Scenario	Tokens/Message	Weekly Cost	Annual Cost
Baseline (target)	5-10K	~$0.15	~$8
Current bloat	130K	~$4/week	~$200
Cost increase	12-26x	26x higher	25x higher

Operational Impact:

• Cost optimization strategy blocked
• Cannot implement weekly context resets
• System unsustainable at current usage rates
• User unable to meet cost targets

---

Steps Taken So Far

1. ✓ Identified context bloat (May 17)
2. ✓ Implemented file-based workflow to reduce conversation context
3. ✓ Attempted container restart (May 19, 8:53 AM)
4. ✓ Verified restart did not resolve issue
5. ✓ Documented metrics and timeline
6. → Escalating for infrastructure investigation

---

Data for Debugging

Session Creation Query:

Session: sess-1778272236516-2bodft
Agent Group: ag-1778272236511-97z4s3 (Prachin)
Container: f15e22cf6a58 (hostname)
Last Active: May 20, 2026, 9:01 AM UTC
Status: active/running

Affected Configuration:

• NanoClaw v2 on Raspberry Pi 5 (4GB RAM)
• Claude Model: Haiku 4.5
• Integration: Telegram bot
• Session Mode: shared

---

Requested Resolution

1. Root Cause Investigation: Why does session context persist after container restart?
2. Database Check: Verify inbound.db/outbound.db cleared on restart
3. Cache Inspection: Check if context cached at higher level
4. Fix: Implement proper context cleanup on container restart
5. Verification: Test that restart actually clears session context

---

Workaround Status

Current Mitigation (File-Based Workflow):

• Using persistent files instead of conversation memory
• Files: meeting-agenda.md, tasks.md, reminders.md, policy-ideas.md
• Reduces context per message, but doesn't address root cause
• Effective only if other users don't inflate context
• Still costs 12-26x baseline

Sustainability: Workaround not sustainable long-term at current cost levels.