Token efficiency
Costs little. Saves more.
Not an estimate, a measurement: we analyzed every single API call to show what an AI memory really costs and why it still pays off.
At a glance
Measured in live operation.
The basis: complete session logs from real-world use with Claude Code and other agents over eight weeks, call by call.
What Hebby costs
It depends on how you work.
Hebby brings the right memories along with each of your requests. The more tool work your agent does per request, the smaller that share gets. The more dialogue, the larger.
3–4 %
Agents that work through many steps in a row on their own. Measured: 4.0 % median for automated sessions, 3.7 % for sessions of 300+ calls.
≈ 5 %
The typical day of questions, code and research. Measured: 5.2 % across all sessions with Hebby active.
up to 10 %
Lots of back-and-forth, little tool work. Every request counts here, and Hebby brings the right context to each one. Measured: 9 to 10 % in dialogue-heavy sessions.
What Hebby saves
The math works out almost every time.
For a single request, your agent performs around 40 work steps on average: searching files, reading logs, checking code. Every step costs tokens. Hebby makes sure many of them are never needed.
Example: a request where your agent would otherwise fall into a known trap
- −15 % 6 research steps vanish: Hebby already knows the answer
- −50 % 20 steps of failed attempt, debugging and correction vanish: Hebby warns about the trap before it blows up
- +4 % Hebby brings the right memories along
- = 61 % stays saved, bottom line
One less research round
Finding a file, re-reading entire log files, checking: easily four to six steps. If Hebby already knows, they disappear.
One less follow-up question
When your agent has to ask, the whole job starts over. Every follow-up question avoided saves an entire request.
One less mistake
Failed attempt, debugging, correction: a repeated mistake often costs more than Hebby costs in a whole month.
Save just two work steps per request and Hebby is in the black.
In our measurements, virtually every request gets memories your agent actually works with. A single research step saved is already far above the threshold. Hebby routinely earns its cost back with room to spare.
Saved twice
Your context files cost more than Hebby.
Everything in CLAUDE.md, AGENTS.md, GEMINI.md and MEMORY.md is sent along with every single call. Even from cache, that costs every time.
A 10,000-token context file costs about as much per call as Hebby does in total. With Hebby it stays nearly empty: knowledge enters the context only when it's needed.
The same goes for built-in memory files and auto-memory. Switch them off and the load drops further. That saving is not even included in our numbers.
Methodology
That's how strictly we measure.
We deliberately price Hebby as high as possible. Every token Hebby causes counts in full as a cost. Everything Hebby saves, we leave out entirely. It doesn't get stricter than that.
- Every call. We analyze the complete session logs, not samples. The cost of each call comes straight from the provider's usage data.
- Every injection. Everything Hebby brings into the context is counted: session start, anticipated memories, queries, connections and even Hebby's own feedback.
- Cache priced correctly. Newly written, cached and cache-read tokens are tracked separately and weighted by the provider's price ratios.
- No offsetting. Research saved, follow-up questions avoided, mistakes prevented, shrunken context files: none of it shows up in our numbers. The real balance is therefore far better.
Even under these toughest conditions, Hebby pays off.
In real-world use, it saves many times what it costs.
Measurement period Aug 25 to Oct 22, 2026 · 143,907 API calls in 3,106 sessions · Hebby metrics from 516 sessions with Hebby fully active since Sep 28, 2026 · Impact from 1,789 rated requests · Weights: uncached 1, cache write 1.25 and 2, cache read 0.1, output 5.
Impact
One hit per request. Every run better.
Hebby's goal is clear: for every request, at least one memory your agent really works with. When that succeeds, every single run is better than without Hebby. We measure whether it does on four levels.
Does the search find the right thing?
Measures the search: does the memory belong to the topic of the request?
Is the knowledge any good?
Measures the quality of the memories. What is incomplete, outdated or missing altogether can never become useful.
Does it arrive at the right moment?
Measures timing: is the good memory in the context exactly when your agent can put it to use? The true yardstick of value.
What comes next?
Memories that anticipate where things are heading. They're ready when the topic develops in that direction.
One used memory per request makes every run better.
Fitting hits are plentiful. Used means only what actually changes a step of your agent. That's why this number is naturally lower than the others, and why it's the strictest quality measure. Hebby hits the goal on virtually every request. And the longer you work with Hebby, the more memories get used.
Cost control
Your numbers. Any time.
Hebby measures for you exactly as it does for us. The built-in statistics in the portal continuously show how many tokens your memory causes, how much of that comes from cache and what share it makes up of your total AI costs.
No black box, no surprises on the bill. Full cost control.
Early Access
Don’t worry, be Hebby.
An agent that remembers. Hebby is now rolling out to our first users. Tell us briefly what you work with.
