Summary

Memory injection Strategy 2: agent has recall_memory / search_memory tools it calls on-demand during execution. Currently only Strategy 1 (context injection) is implemented.

Design Spec Reference

• §7.7 Memory Injection Strategies — Strategy 2: Tool-Based Retrieval

Scope

• recall_memory and search_memory tool implementations
• MemoryInjectionStrategy protocol compliance
• get_tool_definitions() returns tool schemas for LLM
• Integration with ToolRegistry

Research Context

Four independent sources provide implementation blueprints:

1. Memex (arXiv:2603.04257) — CompressExperience/ReadExperience tool pair:

• CompressExperience: archives current context into KV store, replaces working context with structured summary + index table
• ReadExperience: dereferences an index to reinject archived block (deterministic lookup, not semantic search)
• Index-description pairs as compact context representation
• Results: 24.2%→85.6% success, 43% context reduction

2. NeMo Retriever (NVIDIA) — Tool triad blueprint:

• recall_memory(query, top_k) + search_memory(query, keyword_weight) + compact_context()
• Agent decides when to retrieve and when to compress
• 760k tokens/query cost warning — only for complex retrievals

3. Agentic RAG (TDS) — Three agentic patterns:

• Query rewriting: agent reformulates queries for better retrieval
• Sufficiency checking: agent evaluates results and iterates if insufficient
• Dynamic weighting: tool accepts keyword_weight param to shift BM25-vs-vector per query

4. LangChain Context Compression — /compact manual trigger:

• Agent-triggered compaction at clean task boundaries
• Conservative tuning (agents rarely trigger, but choose good moments)

See research/agentic-retrieval-patterns.md and research/memex-indexed-experience-memory.md in project memory.