Usage and Model Performance

When a task becomes expensive, slow, or unstable, start with the usage panel. Memfit shows input, output, cache, context pressure, and speed data so you can tell whether the issue comes from context, model selection, network behavior, or the task itself.

【Image: AIContextToken usage detail panel, labeling Total Input, Total Output, Cache, Pressure, Speed, and Tier Consumption】

Open the Usage Panel

After running a task in the Agent page, open the model/usage entry for the current conversation. A task can include multiple model calls, so compare both a single call and the whole task.

Metric	Meaning	What to do when it looks abnormal
Total Input	Tokens sent to the model.	Narrow the scope, reduce mounted materials, or split the task.
Total Output	Tokens generated by the model.	Ask for tables, summaries, or fixed templates.
Cache	Input tokens reused by provider-side cache.	Check whether context changes too much between turns.
Pressure	How close the context is to the model limit.	Remove irrelevant context or use a larger-context model.
Speed	First-token latency, total time, or throughput.	Compare model, network, proxy, and context size.
Tier Consumption	High-quality, lightweight, and vision model usage.	Check whether simple tasks are using heavy models.

Understand Cache

Cache is not answer caching. It is reuse of repeated input segments by the model provider. It usually works best for stable system prompts, long-term context, and repeated knowledge-base fragments. If each turn rewrites context heavily, cache hits drop.

【Image: Cache hit comparison, showing stable context with high hits and frequently changing context with low hits】

When cache is low:

1. Check whether each turn mounts different files, knowledge bases, or large temporary text.
2. Check whether the task repeatedly asks Agent to restate full background material.
3. Check whether you frequently switch providers or model tiers.
4. Put stable rules in global context and temporary requirements in the current message.

Good cache candidates:

Team coding rules, fixed output format, audit boundaries, long-term project background.

Poor cache candidates:

Changing logs, one-off pasted files, temporary debug output, rewritten task descriptions.

Read the Cache Diagnostic Chart

In the usage details, inspect cache-related charts. If Input is high and Cache is low, the task sends a lot of non-reusable content. If Cache is high but Total Input is still high, the context is reusable but still too large.

【Image: Cache diagnostic chart labeling Input, Cache Hit, Uncached Input, and Pressure】

Chart pattern	Likely cause	Optimization
High Input, low Cache	Too much temporary text or frequent context changes.	Move stable rules to global context and reduce current-turn input.
High Input, high Cache, high Pressure	Stable but oversized context.	Split the task and reduce mounted files or knowledge bases.
High Output	Output request is too open-ended.	Limit fields, length, and format.
High Tier Consumption	Calling mode is too heavy.	Tune model tiers and calling mode in basic configuration.

Analyze Jitter

Jitter means similar tasks vary sharply in latency or quality. Compare at least three similar runs rather than judging one call.

【Image: Model speed jitter analysis chart comparing first-token latency, total time, Pressure, and Input across runs】

Symptom	How to identify it	Fix
First token is slow, generation is normal	First-token latency rises but throughput is fine.	Check provider, proxy, network, or service queueing.
Total time is slow and Output is high	Output token count rises.	Limit output length and ask for conclusion first.
Slowness tracks Pressure	Pressure and latency rise together.	Reduce context, knowledge bases, or mounted files.
Result style drifts	Constraints change between runs.	Fix the output template and move long-term rules into global context.
Simple tasks use high-quality models	Tier Consumption shows heavy model usage.	Check calling mode and use lightweight models for simple tasks.

Reduce Cost

Start with the most effective moves:

1. Specify directories, files, and knowledge-base scope.
2. Require a fixed output format.
3. Split large tasks into smaller tasks.
4. Use lightweight models for simple questions and high-quality models for high-risk reasoning.
5. Keep Review enabled so unrelated tool calls can be caught early.

Example:

Only analyze src/auth. Output a table: risk, evidence file, reason, remediation.
Do not include unrelated content or background explanation.

When to Return to Basic Configuration

If the issue comes from model tiers, pressure thresholds, or Review strategy, return to Basic Configuration and Custom AI Configuration.

Goal	Adjust in basic configuration
Simple tasks are too expensive	Lightweight model, calling mode
Complex tasks are unstable	High-quality model, Review rule
Context often approaches the limit	Pressure token threshold, model context size
Tool usage is too aggressive	Review rule, risk threshold, tool runtime AI review

Next Step

If the metrics show that Agent is not just slow or expensive but going in the wrong direction, read Manual Intervention and Context Editing.