Research

Measured gains from putting memory at inference. Bounded context, lower spend, quality that holds across four domains.

research · production memory

LLMs are stateless. Every call starts cold unless you rebuild context yourself. Full history replay works until sessions get long, then cost grows as O(N²) and attention degrades in the middle of the prompt.

Cortyxia remembers what matters and injects only what this query needs at inference. Context stays bounded, typically 6–12K tokens, turn 1 or turn 50. Quality never takes the hit. Spend does.

80.8%

fewer tokens

governance

10.2×

by question 50

compounding

100%

SWE resolution

vs 73.3% baseline

91.5%

IDE token cut

quality held

What breaks without structured memory

Token inflation. Full-context replay scales linearly per turn and quadratically over a session. In our governance eval, question 50 alone burned 103,972 prompt tokens.

Lost in the middle. The facts are in the prompt, but the model cannot find them. Full-context resolved only 73.3% of SWE-style tasks.

Context silos. Slack, IDE, CRM, and governance tools each keep their own state. Decisions do not travel. Teams re-explain the same policy every session.

cost_shape · session length

Full-contextO(N²)
Sliding windowO(N·W)
Summarizationlossy O(N)
CortyxiaO(N·B)

What changes when memory sits at inference

You do not need the internals to feel the difference. Here is what production teams get.

01

Bounded context

Prompts stop growing with every turn. Cortyxia packs only what this query needs into a fixed budget, so spend stays predictable as sessions get longer.

02

Retrieval temperature

Control how much memory comes back with each call. Focused for everyday answers. Deeper when the work is complex.

03

Memory that travels

Decisions made in one tool stay available in the next. Teams stop re-explaining the same policy every session, and institutional knowledge stops leaking between apps.

04

Quality under load

Less noise in the prompt means the model finds the right fact faster. In our evals, efficiency gains came with matched or better answers, not a quality tradeoff.

Retrieval temperature

LLMs have temperature. So should context retrieval.

Models already let you control how creative a response is. Cortyxia lets you control how much memory comes back with each call. Turn it down for a focused answer. Turn it up when the work is complex and the model needs richer context.

Lower setting

Focused answers

Less context in the prompt. Less noise. Best for policy questions, support, and everyday lookups where you want the right fact, not the full archive.

Higher setting

Deeper context

Richer memory for complex work: coding, investigations, multi-step fixes. More structure when the task needs it, without replaying the entire conversation history.

Context size over a coding session

20 turns

Prompt tokens per turn

startturn 5turn 10turn 15end
Focused Deeper Full history

91.5%

Fewer tokens in a focused coding session versus replaying full history.

~90%

Still cut on the deeper setting, with a bit more context when the task needed it.

100%

Bug-fix success at the highest setting, versus 73% when replaying full history.

Built for production, not demos

Cortyxia sits in the inference path as infrastructure. It remembers across sessions, retrieves what is relevant for the current call, and keeps the prompt lean without asking you to rebuild agents or wire a new integration stack.

proxy or sdk · model-agnostic

01Persistent memory across sessions and tools
02Retrieval temperature: LLM-style dial for context depth
03Fixed token budgets that hold under load
04Drop-in path: one key swap, no rebuild

Token growth stays bounded

Full-context climbs to ~104K tokens by question 50. Cortyxia stays near ~10K. Early turns carry a small retrieval overhead. By question 6 the advantage flips and never looks back.

prompt_tokens · per questiongemini 2.5 flash · n=50
q1
q5
q10
q15
q20
q25
q30
q35
q40
q45
q50
full-context cortyxia10.2× fewer by q50

Evaluated across four domains

Same architecture. Different workloads. Token savings without trading away answer quality.

Enterprise governance

80.8%

Fewer prompt tokens vs full-context on a 50-question session. Quality held. Advantage compounds to 10.2× by question 50.

Gemini 2.5 Flash

IDE coding

91.5%

Token reduction on a 20-turn emulated IDE session with comparable code quality.

Gemini 3.1 Flash-Lite

SWE-style fixes

70%

Fewer tokens, and Cortyxia resolved 100% of tasks vs 73.3% for full-context.

Gemini 3.1 Flash-Lite

LoCoMo (public)

39.8%

Token reduction on a public long-term conversational memory benchmark with comparable quality.

Gemini 2.5 Flash

Better answers, not just cheaper ones

On SWE-style bug fixing, Cortyxia cut tokens by 70% and still beat full-context on resolution. Structured injection helps the model act instead of drowning in replayed history.

On enterprise governance, Cortyxia posted the highest semantic similarity score versus full-context, sliding window, and summarization, while using 5.2× fewer tokens than full-context on average.

swe_resolution · 15 taskstemp 0.9

Cortyxia

100%

Full-context baseline

73.3%

Structured context improved both efficiency and task success

Drop-in infrastructure

Cortyxia runs as a transparent proxy between your app and any chat-completions endpoint. One key swap. No rebuild. No MCP project. SDKs in Python and TypeScript are available when you want programmatic control.

Proxy modePython SDKTypeScript SDKModel-agnostic

Want the full methodology?

We publish measured results and evaluation design. The production system stays proprietary and calibrated per deployment.