Evergreen vs Volatile Context
The ingest-discipline principle for an AI Second Brain: only ingest evergreen context; keep volatile data accessible but out of the brain. The test is a one-year horizon — "will this memory still be useful in a year?" If yes, it belongs (durable background, locked-in decisions, quarterly objectives). If it changes next week — Slack threads, emails, customer records, live CRM data — don't ingest it, or it becomes noise you have to go back and prune every month. Instead, make sure the brain has access to go grab it on demand via Context Routing (project file → wiki → transcripts → the source system, in order). Nate frames this as the split between context (what the business has done, evergreen) and connections (the volatile, changing real data) — the brain is mostly the former.
This is the retrieval-side version of firehose's "signal, not noise / hold what matters" discipline: more ingested data is not more capability, and an always-on autonomous ingest (level 5 of Retrieval Maturity Levels) is exactly where this gate is lost — which is why staying in manual control of what enters the brain is a defensible default. The failure mode to avoid is treating "put everything in the second brain" as the goal; the goal is a store whose every item earns its place.
Claims
- Only ingest evergreen context; the test is "will this be useful in a year?" principle — durable: the store's value is that every item still pays rent later; time-decaying data dilutes it.
- Keep volatile data accessible but not ingested — give the brain access to grab it, don't swallow it. best practice — context: Slack/email/CRM/customer data that changes weekly; "best" is routed live access, because ingesting it creates prune-forever noise while omitting routing loses the answer.
- Ingesting volatile data turns the brain into noise you must prune monthly. observation — the concrete cost of getting the gate wrong.
- Always-on autonomous ingest removes the evergreen/volatile gate — manual control of what enters is a defensible default. best practice — context: level-5 "Brain OS" designs; hands-off ingest is worse than manual precisely when signal quality matters more than convenience.
- Be selective with what pipelines feed the brain — "less is more; only high-signal resources"; don't pump it with everything. best practice — context: a second, independent source (Austin Marchese) reaches the same gate for the inflow of a Self-Improving System — continuous ingestion rivers should carry only high-signal material, or the store fills with noise. Same "signal, not noise" discipline, aimed at the pipelines rather than the one-time ingest.
Related
- AI Second Brain — this is the discipline for what the brain is allowed to hold.
- Retrieval Maturity Levels — level 5's always-on ingest is where this gate is lost.
- Context Routing — how volatile data stays reachable without being ingested.
- Query-Shaped Storage — the complementary question: shape what you do store by the query.
- Self-Improving System — the inflow pipelines this discipline governs: feed the rivers only high-signal, evergreen material.
- Distillate: Every Level of a Claude Second Brain Explained
- Distillate: How to Build a Self-Improving System with Claude Code — "less is more, only high-signal resources," applied to continuous ingestion pipelines.
Linked from
- Agent Loop
- AI Second Brain
- Capture-Organize-Distill-Express (CODE)
- Capture-Storage-Retrieval Pipeline
- Compiled Knowledge Base
- Context Compression
- Context Smart Zone
- Context Substrate
- Every Level of a Claude Second Brain Explained
- How AI Agents Search Their Memory — Hybrid Retrieval, in Practice (OpenClaw)
- How to Build a Self-Improving System with Claude Code
- I Gave Claude Code a Permanent Memory
- Prompt Caching
- Recency-Grounded Research
- Retrieval Maturity Levels
- Search-Then-Get
- Second Brain Explained for Engineers and Knowledge Workers
- Self-Improving System
- Skill-Driven Loop Development
- Loop Engineering, Illustrated: Triggers, Skills, Verification, Memory