Athena’s memory system combines structured state (Memory Bank) with semantic recall (VectorRAG) to create a compound learning loop where each session starts smarter than the last.
Think of it as the difference between:
Vector DB: “Where did I see something about X?” (search)
Memory Bank: “What am I working on right now?” (state)
When the total exceeds 15K tokens, activeContext.md auto-compacts—merging older session summaries into shorter entries until the budget is back under 10K.
Problem: Learnings written to files (e.g., User_Profile_Core.md) become passive documentation. The AI doesn’t read them unless explicitly prompted, causing the same mistakes to repeat.
Solution: Active Injection — Force-feed critical constraints into the terminal during boot.Key Scripts:
boot_knowledge.py: Extracts and prints constraints
index_workspace.py: Rebuilds TAG_INDEX.md and PROTOCOL_SUMMARIES.md on shutdown
Query: “Should I accept this commission-based partnership where the agent takes no risk?”
Generic AI (No Memory)
Athena (With VectorRAG)
Output: “Commission-based partnerships can be effective for motivation. Ensure you have clear contracts. Pros include low fixed costs, while cons include potential short-term focus by the agent.”Verdict: Safe but generic. Balanced pros/cons list but lacks strategic conviction.
Retrieval:
Protocol 33: Principal-Agent Problem (Theory)
CS044_ILP_Trust_Arbitrage (Similar bad deal case study)
Protocol 121: Amoral Realism (Framework)
Output: “Stop. This structure matches the Principal-Agent Problem we analyzed in the ILP Trust Arbitrage Case (CS044). The counterparty has zero downside risk while you bear all liability. This mirrors the ‘Tuition Trap’ mechanic we identified in Dec 2024. Recommendation: Reject the deal unless they accept a clawback clause (shared risk) per Protocol 33.”Verdict: Contextualized & decisive. Uses past organizational memory to prevent recurring errors.