Bug Report: Session Compaction System Not Working

Date: 2026-02-23
Reporter: Ian Baistow-Davies ([email protected])
OpenClaw Version: 2026.2.22-2
Platform: Windows 11 (10.0.26200), Node 22.13.1
Issue Type: Critical - Memory/Performance

Summary

Session compaction is not functioning despite being configured with mode: "safeguard" and contextTokens: 32000. Active Telegram sessions are exceeding the 32k token limit (reaching 126-159%) within minutes of activity, indicating the compaction system is not triggering at the expected 80% threshold (25.6k tokens).

Expected Behavior

According to OpenClaw documentation:

1. Compaction should trigger at 80% capacity:

   • With 32k context: compaction at ~25.6k tokens
   • Mode "safeguard": Conservative automatic compaction

2. Context pruning should remove old content:

   • TTL: 15 minutes
   • Mode: "cache-ttl"
   • Old messages should be pruned after idle period

3. Sessions should not exceed configured limit:

   • contextTokens: 32000
   • Sessions should stay within or compact before reaching limit

Actual Behavior

Timeline of Observed Issues

Initial State (Before Update)

• Date: 2026-02-23 02:00 AM
• Telegram session: 51k/32k tokens (159%)
• Age: Only 22 minutes old
• Issue: Uncontrolled growth, compaction never triggered

After Update to 2026.2.22-2 (First Check)

• Date: 2026-02-23 04:17 AM
• Telegram session: 21k/32k tokens (65%)
• Status: Appeared fixed - session properly managed
• Conclusion: Update seemed to include compaction fixes

After Force Reinstall (Current State)

• Date: 2026-02-23 04:30 AM
• Telegram session: 40k/32k tokens (126%)
• Age: Only 13 minutes
• Issue: Rapid growth resumed - 3.1k tokens/minute growth rate

Key Observations

1. Compaction never triggers:

   • Sessions routinely exceed 32k limit
   • No evidence of compaction at 80% threshold (25.6k)
   • Sessions grow unbounded until system intervention

2. Context pruning ineffective:

   • Sessions retain full history despite 15m TTL
   • Stale cron sessions (2+ hours old) still contain 74k tokens
   • No automatic cleanup of old context

3. Growth rate unsustainable:

   • Telegram session: 40k tokens in 13 minutes = ~3.1k tokens/min
   • At this rate: 93k tokens/hour
   • Cost impact: Significant (prompt caching only 49% effective)

Configuration

File: C:\Users\ianba\.openclaw\openclaw.json

{
  "agents": {
    "defaults": {
      "contextTokens": 32000,
      "contextPruning": {
        "mode": "cache-ttl",
        "ttl": "15m"
      },
      "compaction": {
        "mode": "safeguard"
      },
      "maxConcurrent": 1
    }
  }
}

Session Configuration:

• Default model: google/gemini-2.5-flash
• Context window: 32k tokens
• Session target: "main" (persistent sessions)
• DM scope: "per-channel-peer"

Evidence

Current Session Status (2026-02-23 04:30 AM)

Sessions
┌──────────────────────────────────────────────┬────────┬─────────┬──────────────────┬────────────────────────────────┐
│ Key                                          │ Kind   │ Age     │ Model            │ Tokens                         │
├──────────────────────────────────────────────┼────────┼─────────┼──────────────────┼────────────────────────────────┤
│ agent:main:telegram:direct:7249…             │ direct │ 13m ago │ gemini-2.5-flash │ 40k/32k (126%) · 🗄️49% cached │
│ agent:main:main                              │ direct │ 13m ago │ gemini-2.5-flash │ 28k/32k (87%)                  │
│ agent:main:cron:3db4f6c4-8692-4…             │ direct │ 2h ago  │ gemini-2.5-flash │ 74k/32k (232%) · 🗄️2% cached  │
└──────────────────────────────────────────────┴────────┴─────────┴──────────────────┴────────────────────────────────┘

Growth Pattern Analysis

Expected behavior: Should have compacted at 25.6k (80%)
Actual behavior: Grew from 21k → 40k without compaction

Impact Assessment

Performance Impact

• Memory usage: Sessions accumulating unbounded context
• API costs: Higher token usage on every request
• Response latency: Larger context = slower processing

Cost Impact (Estimated Annual)

Telegram session alone:

• Current: 40k tokens/request at ~12 requests/hour
• Expected: 25k tokens/request (with proper compaction)
• Excess: 15k tokens × 12 requests × 24 hours × 365 days
• Additional cost: ~$789/year (at $0.0001/1k input tokens)

All bloated sessions combined:

• Estimated additional cost: $2,000-3,000/year

Related Issues

1. Isolated Session Mode Bug (Issue #24101)

• Status: Previously reported 2026-02-23
• Issue: Isolated cron sessions accumulate context despite being marked "isolated"
• Evidence: pairing-watch session at 74k tokens after 22 hours
• Relation: Both issues suggest compaction system fundamentally broken

2. Memory System Module Import Failure (Resolved)

• Issue: Hash mismatch in module imports after npm update
• Error: Cannot find module 'manager-DRT2R_zd.js'
• Resolution: Force reinstall with npm install -g openclaw@latest --force
• Status: ✅ Fixed - Memory system now operational

Reproduction Steps

1. Install OpenClaw 2026.2.22-2
2. Configure session with:

   {
     "contextTokens": 32000,
     "compaction": { "mode": "safeguard" },
     "contextPruning": { "mode": "cache-ttl", "ttl": "15m" }
   }

3. Start Telegram conversation with agent
4. Send 10-15 messages with tool calls over 10-15 minutes
5. Check openclaw status → session will exceed 32k tokens
6. Expected: Compaction at ~25.6k tokens
7. Actual: No compaction, session grows to 40k+ tokens

Attempted Solutions

1. ✅ Updated to 2026.2.22-2

• Result: Temporarily appeared to fix (session at 21k)
• Outcome: Issue returned after reinstall (session at 40k)

2. ❌ Attempted "aggressive" compaction mode

• Error: "Invalid input" - mode not recognized
• Available modes: Unclear (documentation needed)

3. ✅ Force reinstall to fix module hash mismatch

• Result: Memory system fixed, but compaction still broken

4. ⏳ Pending: Restart gateway to clear stale sessions

• Status: Not yet attempted for this test

System Information

Environment:

OS: Windows 11 (10.0.26200 x64)
Node: 22.13.1
OpenClaw: 2026.2.22-2
Gateway: ws://127.0.0.1:18789 (local loopback)
Tailscale: serve · ianlaptop.tail713286.ts.net

Agents: 5 total (main, friends, rory-a, rory-b, switcher)
Memory: 42 files · 90 chunks · vector ready · fts unavailable
Channels: Telegram (2 accounts enabled)

Requested Actions

1. Investigate compaction system:

   • Why does "safeguard" mode not trigger at 80% threshold?
   • Are there logs showing compaction attempts or failures?
   • Is there a configuration issue preventing compaction?

2. Document valid compaction modes:

   • What modes are available? (safeguard, aggressive, off, on?)
   • What are the differences between modes?
   • When should each mode be used?

3. Fix or document workaround:

   • If compaction is broken, provide hotfix or manual compaction command
   • If configuration issue, provide correct configuration example
   • If by design, document how to prevent unbounded growth

4. Context pruning investigation:

   • Why do 2+ hour old sessions still retain full context?
   • Is "cache-ttl" mode working as documented?
   • Should TTL be shorter? Different mode?

Additional Notes

Workaround Currently Using

Manual session cleanup:

1. Archive bloated sessions periodically
2. Delete session files to force restart
3. Monitor with custom PowerShell scripts

Script: C:\Users\ianba\Rorys_play_area\scripts\backup\archive-bloated-sessions.ps1

Frequency: Weekly or when sessions exceed 500KB

Effectiveness: Temporary - sessions rebloat within hours/days

Logs

Memory system log snippet:

[31m[memory][39m [33mfts unavailable: no such module: fts5[39m

(Note: FTS unavailability is separate issue, not related to compaction)

Session store locations:

• Main: C:\Users\ianba\.openclaw\agents\main\sessions\sessions.json
• Friends: C:\Users\ianba\.openclaw\agents\friends\sessions\sessions.json

Session file sizes:

• Telegram session: ~1.1 MB (before cleanup)
• Largest cleaned: 4.4 MB
• Total cleaned: 17.49 MB across 14 sessions

Priority Justification

Priority: CRITICAL

Reasons:

1. Cost impact: $2-3k/year in excess API costs
2. Performance degradation: Slower responses, higher memory usage
3. System stability: Unbounded growth will eventually cause crashes
4. User experience: Degraded performance for all users affected
5. Scope: Appears to affect all session types (main, isolated, telegram)

User Impact:

• All OpenClaw users with active conversations
• Especially high-frequency users (many messages/day)
• Both isolated and persistent session modes affected

Related Files

Configuration: C:\Users\ianba\.openclaw\openclaw.json
Session storage: C:\Users\ianba\.openclaw\agents\main\sessions\
Archive: C:\Users\ianba\Rorys_play_area\archive\sessions-2026-02-23-0411\
Cleanup script: C:\Users\ianba\Rorys_play_area\scripts\backup\archive-bloated-sessions.ps1
Previous bug report: C:\Users\ianba\Rorys_play_area\BUG-REPORT-ISOLATED-MODE.md (Issue #24101)

Report Generated: 2026-02-23 04:35 AM GMT
Contact: [email protected]
Telegram: @IDMcrnYork (User ID: 7249188495)