# Stop wasting tokens on shared selective persistent memory for agentic LLMs > Source: > Published: 2026-09-16 17:08:00+00:00 # Stop wasting tokens on shared selective persistent memory for agentic LLMs Agentic systems that write code via multi-turn tool use have a massive context flaw because every new session wipes the slate clean. You lose the data schemas, domain constraints, and specific configuration choices that actually made the previous session work. Just dumping the whole conversation history into the next prompt is a bad move—it's a token sink and the irrelevant noise actually kills the generation quality. The fix is shared selective persistent memory. Instead of saving everything, this architecture filters for four specific types of reusable context: task specifications, data schemas, tool-use patterns, and configuration choices. ## How this memory architecture actually works The goal here is to keep the "gold" and toss the "trash." When an agent completes a task, the system doesn't just archive the log; it extracts the structural knowledge. - **Task Specifications:** The core requirements and constraints that define the goal. - **Data Schemas:** The actual structure of the data the agent is interacting with, so it doesn't have to re-learn the API or DB schema every single time. - **Tool-Use Patterns:** Which tools worked for which specific problem, reducing the trial-and-error loop in the next session. - **Configuration Choices:** Specific settings or environment variables that were settled upon during the first few turns. ## Getting the system to actually retain the right data To implement this, you can't just use a basic vector DB retrieval because that often pulls in "similar" but irrelevant conversational filler. You need a selective layer that categorizes the memory before storage. 1. **Extraction Phase:** After a successful tool call or task completion, a separate "summarizer" pass identifies if any of the four categories above were established. 2. **Storage Phase:** These are stored as structured metadata rather than raw chat logs. 3. **Injection Phase:** When a new session starts, the system checks the task ID or user profile and injects only the relevant schemas and specs into the system prompt. ## Is it actually worth the overhead? If you're running a simple chatbot, this is overkill. But if you're building a coding agent that manages a complex codebase over several days, it's a necessity. The trade-off is a bit of extra latency during the "cleanup" phase after a session ends, but you save a massive amount on input tokens for every subsequent prompt. More importantly, you stop seeing the agent hallucinate a different data schema than the one it used ten minutes ago. The real win here is the "selective" part. By ignoring the conversational fluff and only persisting the technical constraints, you keep the context window clean and the model's attention focused on the current task rather than a transcript of how it got there. [Next NVLink 6 handles failures so AI factories don't stop →](https://promptcube3.com/en/threads/9431/) ## All Replies （0） No replies yet — be the first!