Memory System Improvements - Summary

Sync Note (2026-03-10)

This summary is synchronized with the main branch implementation. TF-IDF/context-aware retrieval is planned, not merged yet.

Implemented

• Accurate token counting with tiktoken in memory injection.
• Facts are injected into <memory> prompt content.
• Facts are ordered by confidence and bounded by max_injection_tokens.

Planned (Not Yet Merged)

• TF-IDF cosine similarity recall based on recent conversation context.
• current_context parameter for format_memory_for_injection.
• Weighted ranking (similarity + confidence).
• Runtime extraction/injection flow for context-aware fact selection.

Why This Sync Was Needed

Earlier docs described TF-IDF behavior as already implemented, which did not match code in main. This mismatch is tracked in issue #1059.

Current API Shape

def format_memory_for_injection(memory_data: dict[str, Any], max_tokens: int = 2000) -> str:

No current_context argument is currently available in main.

Verification Pointers

• Implementation: packages/harness/deerflow/agents/memory/prompt.py
• Prompt assembly: packages/harness/deerflow/agents/lead_agent/prompt.py
• Regression tests: backend/tests/test_memory_prompt_injection.py