Feature Request: Post-Compaction Recovery Turn

Problem

When auto-compaction completes, the agent loses context about what it was doing. The existing memoryFlush (pre-compaction) helps save state, but there's no mechanism to recover that context after compaction.

Current flow:

1. ✅ Pre-compaction: memoryFlush saves durable memories
2. ✅ Compaction: summary generated, old messages removed
3. ❌ Post-compaction: agent continues with reduced context, doesn't know to read saved memories

User impact: After compaction, users must re-explain context or the agent "forgets" ongoing tasks.

Development Process

Attempt 1: Hook into existing compaction flow

Approach: Add recovery logic directly in agent-runner-memory.ts after compaction detection
Problem: The compaction flag is checked but there was no hook point to inject a new turn
Outcome: ❌ Needed a cleaner abstraction

Attempt 2: Create dedicated post-compaction module

Approach: New post-compaction.ts module with:

• buildPostCompactionTurn() - constructs the recovery message
• shouldRunPostCompaction() - checks config and compaction state
• Integration in agent-runner.ts after compaction completes

Outcome: ✅ Clean separation, testable, works

Attempt 3: Config schema design

Approach: Mirror the existing memoryFlush pattern
Options considered:

• compaction.postCompaction - too nested
• memory.postCompaction - consistent with memory features
• recoveryPrompt - less discoverable

Outcome: ✅ Chose memory.postCompaction.enabled + memory.postCompaction.prompt

Testing

Test 1: Basic trigger

• Filled context to 151k tokens to force compaction
• ✅ Post-compaction turn injected after compaction completed

Test 2: Memory recovery with marker

• Added BANANA_TEST_7X9K2 to memory/2026-01-28.md
• Triggered compaction (151k → 105k tokens)
• ✅ Agent read memory file and reported the marker correctly

Test 3: Multiple compaction cycles

• Continued session, triggered second compaction
• New marker: BANANA_KEY_8472
• ✅ Recovery worked again across cycles

Known limitation found during testing

• contextTokens config doesn't override model's native context window
• This is a separate issue, not related to this feature

Solution

New memory.postCompaction config:

{
  "memory": {
    "postCompaction": {
      "enabled": true,
      "prompt": "Compaction completed. Read memory/YYYY-MM-DD.md to recover context..."
    }
  }
}

How it works:

1. After compaction completes, check if postCompaction.enabled
2. If enabled, inject a system turn with the configured prompt
3. Agent reads memory files and recovers context
4. User continues seamlessly without re-explaining

Files Changed

Note on Contribution Guidelines

Per CONTRIBUTING.md, new features should start with a Discussion. This issue was created alongside the implementation as a proof-of-concept. Happy to move to Discussion format if preferred by maintainers.