The Context Ledger · 2026 baseline study

Your AI coding agents keep re-learning your codebase. Here's the bill.

We measured 6,104 real coding-agent sessions from 309 developers using exact per-request API accounting — the token counts the vendors themselves report, reconciled to 100% of billed volume. Not a survey, not a simulation of synthetic tasks: real work, exactly priced.

$ npm install -g @membridgeai/membridge $ membridge audit Or download the app

The audit ships with MemBridge: install it, then membridge audit prints your number. It reads the session logs your agents already keep. Nothing leaves your machine.

6,104sessions analyzed
309developers
211production repos
76.3Btokens, billed volume
$56,666spend, exactly accounted
Finding 1

The median developer burns 31% of agent tokens on re-orientationmeasured

Re-orientation is your agent re-reading, re-searching, and re-deriving context in a codebase it has already worked in. Across the 174 developers with five or more sessions, the median share of billed tokens spent this way is 31.2%, and nine in ten developers sit above 16%. At the corpus level, re-orientation accounts for $14,524 of the $56,666 measured — 26% of everything spent.

0%re-orientation share of billed tokens, per developer63%

We measured ourselves first. Our own repositories sit near the population's 10th percentile — more efficient than roughly nine in ten developers we measured — and we still burn 13.5%. We built MemBridge because even the efficient end of this chart is paying a permanent tax.

In repos with history, 53% of file reads repeat an earlier session's workmeasured

Median across the 96 repositories with enough session history to measure, ranging as high as 81%. Your agent's breakthroughs from this morning are invisible to the session you open after lunch, and everything your teammate's agent figured out is gone. The waste is worst exactly where teams work together.

And it never goes away on its ownmeasured

Orientation share is essentially flat with repo experience: a developer's twentieth session in a repo re-orients nearly as much as their first. Agents do not learn their way around your codebase. Without a memory layer, the tax is permanent.

Median re-orientation % of billed volume by session order within (developer, repo).

What's recoverable

Shared memory recovers 6–8% of total agent spend, net of its own cost

We modeled savings the hard way: only orientation that is measurably redundant counts, every token is priced at actual cache rates, and the memory layer is charged for the context it injects. The result is a band, not a point, because the model's one assumption — the replacement cost of injected memory, r — is stated rather than hidden.

net recoverable 5.8–7.5%
gross 8.3%
ceiling 24.3% total spend
TierOn this corpusOf total spend
Ceiling: all re-orientation eliminated (not achievable — first visits are real work)$13,77624.3%
Gross: all measurably redundant orientation eliminated$4,6928.3%
Net, memory cost r = 10% stated assumption$4,2237.5%
Net, r = 20% stated assumption$3,7536.6%
Net, r = 30% stated assumption$3,2845.8%

The number survived its own stress test. This band was first computed on 20 developers. Recomputed on a corpus fourteen times larger, it reproduced unchanged — even though orientation share doubled between the two, because the increase is first-time reading no memory can recover, and the model knows the difference. We'd rather publish a number that holds than a bigger one that doesn't.

What that means at scale

For a developer using agents heavily, that band is $1,900–$5,200 a year in tokens alone. For a 50-person agent-first engineering org paying three or four vendors at once, it's a real line item. Extrapolated to the roughly $10B enterprise coding-agent market at the measured 26% orientation rate, the industry is burning on the order of $2.5B a year re-deriving what its agents already knewextrapolation — measured rate times cited market size, stated exactly as that.

And the dollars are the smaller half. The median session spends its first several turns re-orienting before it touches a file: minutes of a developer watching a spinner, several times a day. MemBridge collapses that to one or two. Tokens are what the audit prints; time is what you feel by Wednesday.

Method & data

How the numbers are made

Every API request reports its exact context size: input, cache-creation, and cache-read tokens. The growth between consecutive requests, minus the model's own output, is exactly what the agent's tool results injected between them. Summing gives total billed volume; each block's cost is its size times its residency, priced at published rates with cache multipliers. The ledger reconciles to 100% of billed volume by construction, and sessions that fail reconciliation are excluded and counted, never approximated.

Classification is deliberately conservative: only unambiguous read-only inspection counts as orientation, so every waste figure here is a floor.

Where the data comes from

Real sessions, no telemetry of ours: the SWE-chat research corpus of consenting open-source developers (Stanford SALT, used under ODC-BY; Baumann et al., 2026, arXiv:2604.20779), public Entire checkpoint data from open-source repositories, and our own transcripts. All figures published here are aggregates under anonymous labels.

Your number is sitting in your logs right now.

The audit is part of MemBridge: install it, run membridge audit, and it reads the session files your agents already keep on disk and prints what re-orientation cost you this year. One command, local only. Then MemBridge — one shared memory across the AI coding tools you already use — makes the number go down.

$ npm install -g @membridgeai/membridge && membridge audit