rushikeshsakharleofficial Smart Memory Management
Guide for agent memory. Goal = avoid context rot, save tokens, keep reasoning high.
Memory Architecture (4 Tiers)
- Short-Term (Working): Last 5-10 turns + active task state. In context window.
- Episodic: Past interactions + tool results. Vector DB with recency decay.
- Semantic: Knowledge base, facts, preferences (e.g.,
MEMORY.md, SQL). - Procedural: Core skills, persona, rules (system prompts).
Context Optimization
- Reduction: Summarize history. Keep last N turns full, compress older turns recursively.
- Offloading: Move large tool outputs (HTML, big JSON) to disk. Keep structured pointer in context.
- Selection: RAG (Retrieval-Augmented Generation). Pull only relevant snippets.
- Isolation: Split tasks to sub-agents. Prevent context drowning.
Advanced Tactics
- Strategic Forgetting: Prune noise (false starts, errors) after task done.
- Structured Note-taking: Use scratchpad (
NOTES.md) for long-horizon state tracking. - Tool Condensation: Tools return pre-processed data (row counts, keys), not raw payloads.
- Prompt Caching: Cache stable system prompts + large reference docs.
Implementation Limits
- Threshold Compaction: Trigger summarize/offload when context hits 80-90%.
- Recoverability: Test if agent can find needle-in-haystack after compression.