Skip to content
MemStrata MemStrata

First Time Seeing Savings

After your first AI coding session with MemStrata active, the dashboard will populate with your real numbers. Here’s what you’ll see and what it means.

When do numbers first appear?

Section titled “When do numbers first appear?”

The dashboard needs at least one complete AI request/response cycle to compute metrics. This means:

  • Money tab — updates after your first prompt where MemStrata compressed context. Usually visible within 2–3 coding turns.
  • Retrieval recall and precision — computed after each turn. First values appear after turn 1.
  • 24h edit survival — requires a 24-hour window. Empty on day 0; fills in after your first day of usage.
  • Tokens vs. RAG baseline — appears after your first session once MemStrata has enough context to compute the cohort comparison.

Reading the Money tab

Section titled “Reading the Money tab”
┌─────────────────────────────────────────────────┐
│ Money tab                                        │
│                                                  │
│  Saved this cycle: $14.28                        │
│  vs. $14.99/mo CAD Pro subscription                     │
│                                                  │
│  Daily rate: $0.54 (averaged over active days)   │
│  Projected (30 days): $16.20 ─── On track        │
└─────────────────────────────────────────────────┘

Saved this cycle — total dollar value of token savings since the start of your current billing cycle. Computed as:

tokens_saved × your_provider_rate_per_token

Daily rate — savings divided by number of active days (days where MemStrata processed at least one request).

Projected (30 days) — extrapolation of the daily rate to the end of your billing cycle. Used to predict whether you’ll exceed your subscription cost.

Status indicator:

  • On track — projected savings ≥ subscription cost. No credit expected.
  • Credit likely — projected savings < subscription cost. MemStrata will issue a Stripe credit note at cycle end.

The four metrics explained

Section titled “The four metrics explained”

Tokens vs. RAG baseline

Section titled “Tokens vs. RAG baseline”

The percentage reduction in input tokens compared to what a naive retrieval-augmented generation (RAG) pipeline would have sent. The baseline is computed from the median token count of your cohort (users with similar codebase sizes and tool configurations) using a standard BM25 + embedding top-5 retrieval.

Median: 42%. Your number will vary depending on codebase density, how much your AI requests reference distant parts of the codebase, and which integration path you’re using.

Retrieval recall

Section titled “Retrieval recall”

After each turn, MemStrata parses the AI’s response for referenced symbols, files, and concepts. Recall = fraction of those references that were in the context MemStrata provided.

Median: 87%. A recall drop below ~70% means MemStrata is missing relevant context — this shows up in the dashboard as a flag. Common cause: a part of the codebase that hasn’t been indexed yet (run memstrata index to refresh).

Retrieval precision

Section titled “Retrieval precision”

Precision = fraction of entities MemStrata included in context that the AI actually referenced. High precision means focused context; low precision means padding.

Median: 64%. Lower precision inflates your token count. The dashboard surfaces this metric to help you tune context budget settings.

24h edit survival

Section titled “24h edit survival”

Fraction of AI-suggested edits that are still present in the repository 24 hours after being applied. A proxy for suggestion quality — bad suggestions tend to get reverted within hours.

Median: 73%. This metric is read-only; MemStrata does not gate on it. It’s surfaced so you can see whether suggestion quality correlates with specific tools or context configurations.

What if my numbers look wrong?

Section titled “What if my numbers look wrong?”

Numbers are zero or very low:

  • Verify MemStrata is running: memstrata status
  • Confirm your AI tool is routing through the harness (check the harness log at ~/.memstrata/harness.log)
  • For MCP integrations: verify the MCP server is listed in your tool’s config

Recall is very low (<50%):

  • Run memstrata index to refresh the codebase graph
  • Check that your project root is correctly configured: memstrata config show

Savings look unexpectedly high:

  • This is possible if your codebase has many large files that MemStrata is aggressively compressing. The absolute savings are real; verify by checking a specific session in the Sessions tab.

See Troubleshooting for more diagnostic steps.