Context

Build the memory retrieval pipeline that fetches relevant memories and injects them into the agent's context before LLM calls. This is what makes memory actually useful — without injection, stored memories are never consulted.

Acceptance Criteria

Retrieval Pipeline

• Retrieve relevant memories based on current task/query context
• Rank results by relevance and recency (configurable weighting)
• Token budget enforcement for memory injection (don't exceed context window)
• Graceful degradation when too many relevant memories exist (truncate/summarize)

Context Injection

• Inject retrieved memories into system prompt or conversation history
• Configurable injection point (system prompt vs user context vs both)
• Clear formatting of injected memories (agent can distinguish memory from instruction)

Shared Knowledge

• Retrieve from shared knowledge base (cross-agent memories)
• Merge personal and shared memories with appropriate priority

Testing

• Unit tests for retrieval and ranking logic (>80% coverage)
• Integration test demonstrating agent behavior change when memories are available vs absent

Dependencies

• Depends on SQLite persistence (Implement company templates loading and interactive builder CLI #38)
• Depends on Mem0 evaluation (Evaluate memory layer candidates: Mem0, Zep, Letta, Cognee, custom #39)

Design Spec Reference

• Section 7 — Memory System