An agent starts paying before it starts working. It loads instructions, project state, decisions, examples, safety rules, tool descriptions, and historical context before the first useful action. That standing context is the preamble tax: token, latency, and attention cost paid at the beginning of a session.
The problem is not that every preamble is waste. The problem is that teams rarely know which part is necessary, which part is stale, which part was actually served, or whether their measurement includes everything the agent was required to read.
One preamble, two different mechanisms
Standing context reaches an agent in two ways.
Automatic context
The runtime injects automatic context. A system prompt, repository instruction file, tool schema, or memory payload appears in the session without the agent first deciding to fetch it.
This cost is visible. A trace can show that the content was served, and a measurement tool can count it.
Mandated context
Mandated context is different. One instruction tells the agent to read another source before it acts:
- Read the engineering process.
- Load the current product decision.
- Inspect the deployment topology.
- Review the task specification.
The requirement is real, but the read may not be automatic. Unless the workflow records it, an observer cannot tell whether it happened. A measurement that counts only startup injection excludes the context the agent must fetch afterward.
The automatic half announces itself. The mandated half can remain invisible.
Correct arithmetic can answer the wrong question
We learned this by measuring the standing context in our own Northsun (formerly Mnemix) repository. The first pass produced a reproducible result from an explicitly defined input set. The calculation was correct.
The input set was incomplete.
It covered automatically loaded sources but missed mandatory follow-on reads named inside those sources. The instrument answered “what loads automatically?” while the intended question was “what must a compliant agent consume before useful work?”
Nothing about clean arithmetic reveals an omitted input. A system can calculate the wrong scope perfectly.
We retained the original measurement beside its correction because deletion would hide the failure mechanism. The important repair was not a more persuasive number. It was making the measurement boundary visible on the face of every report so another reviewer could challenge its completeness.
Why giant instruction files keep growing
Teams often respond to agent amnesia by adding more startup text:
- An agent misses a decision.
- The decision is copied into a mandatory instruction file.
- The file becomes longer and more expensive to load.
- Some copied decisions become stale.
- More explanation is added to resolve the conflict.
The preamble grows because the file is being used as storage, routing, authority, and delivery at the same time.
This can work at small scale. It becomes fragile when several agents, repositories, environments, and sessions need different slices of context. Every consumer receives the whole snapshot, even when only a few facts are relevant.
Audit your preamble tax without inventing a benchmark
You do not need a public benchmark or a perfect tokenizer to find the structural problem. Start with a scope audit.
1. Inventory automatic sources
List every source the runtime injects before the first user task. Record the owner, loading mechanism, freshness expectation, and evidence that it was actually served.
2. Follow every mandatory-read edge
Search the automatic sources for instructions such as “must read,” “start here,” “before implementation,” or “refer to.” Follow those references recursively. A mandatory source is part of the operational preamble even when it is not injected automatically.
3. Separate stable doctrine from live state
Stable rules and volatile facts should not share the same refresh model. Branch names, deployment state, current owners, versions, pricing, and endpoints can become wrong quickly. A permanent preamble is a dangerous home for them.
4. Record the boundary
Every measurement should state exactly what it counted, what it excluded, and why. “Total context” is not a useful label when the instrument cannot prove completeness.
5. Test the failure mode
Remove a mandatory source from a fixture or add an undiscovered mandatory-read edge. The audit should fail. A gate never observed failing is not yet a proved gate.
The goal is selective, governed context
The alternative to a giant preamble is not zero context. It is context selected for the current subject, moment, tenant, and policy—with enough evidence to explain why it was served.
That requires more than retrieval:
- Identity: whose context is eligible?
- Authority: which source wins when statements conflict?
- Freshness: was this true for the moment being reconstructed?
- Provenance: where did the item come from?
- Policy: may it be used in this session?
- Salience: is it useful now?
- Observability: can the system show what it served and what it omitted?
Northsun is the memory and enrichment layer for AI agents. Its public API and MCP surface provide current integration paths, while broader governed-context controls remain explicitly state-labeled in the documentation. Northsun does not currently claim to delete instruction files automatically or guarantee token-cost reductions.
The practical direction is still clear: stop treating every past instruction as permanent startup context. Route the smallest trustworthy context that helps the agent act, and preserve the evidence needed to challenge that choice later.
Continue with Core concepts, Context rot, and Evidence references.