01
ACTION
From conversation to action
Tool use and coding agents connected reasoning to files, terminals and tests, letting results be checked against the real environment.
Reasoning meets realityEVOLUTION / AN ENGINEERING DIRECTION
From doing the work
to improving how it gets done.
Models improve over time. Memory, skills, tools and workflows can improve too. EvoForge is developing ways to carry useful experience forward, with evaluation and approval keeping each change accountable and under your control.
01 / THE TRAJECTORYACTION → PERSISTENCE → LEARNING
Agent evolution has a history.
Each step opens a new possibility.
Tool use, persistent memory, reusable skills and evaluation loops form several parallel paths toward agents that learn from real work.
02
PERSISTENCE · OPENCLAW
An agent gets a place to live
Resident agents such as OpenClaw brought channels, memory and extensible skills together, allowing work and context to continue across time.
Work gains continuity03
PROCEDURAL MEMORY · SKILLS
Experience becomes procedure
Skills package instructions, references and scripts as reusable methods. Review and maintenance keep those methods relevant as the system grows.
Experience becomes reusable04
LEARNING LOOPS · HERMES
Candidates enter evaluation
Systems such as Hermes connect persistent memory with skill revision. Self-evolution research extends evaluation to prompts, skills, tool descriptions and code.
Change becomes measurable05
METACOGNITION · EVOFORGE
The system learns where to improve
EvoForge is exploring a metacognitive framework that recognizes recurring gaps, identifies whether the next improvement belongs in knowledge, procedure, tools or models, and evaluates the result.
The system reasons about its own growth02 / THE KERNELFOUR LAYERS · ONE CODEBASE
One integrated system,
from context to action.
Written in Rust. Replaceable models connect at the top; your computer, apps and tools form the working environment below. Context, memory, permissions and orchestration connect the two.
FIG.01 / KERNEL · GENERAL ARRANGEMENTSECTION VIEW
Context engine
Attachments and tool results are organized within a context budget, helping long tasks stay focused and continuous.
Built-in tools
Ninety-eight tools cover document processing, code editing, browser and desktop interaction, terminal work and media creation.
Memory
Facts, preferences and project context are kept in the right scope, with expiry and recall controls that keep future work relevant.
Permissions and sandbox
Risk is assessed using the target, data sensitivity, recovery options and external impact, guided by the autonomy level you choose.
Orchestration
A lead agent or shared task board coordinates specialists, long-running work and scheduled tasks, with progress gathered in one place.
Model layer
Connect cloud or local models and choose the one that best fits each task, with provider health and credentials managed separately.
03 / MODEL INDEPENDENCEINTELLIGENCE ABOVE THE PROVIDER
Models change.
What you have built does not reset.
Foundation models supply the reasoning. EvoForge keeps the memory, skills, specialists and evidence that make that reasoning useful in your world, then carries them to whichever model you choose next.
FIG.02 / MODEL-INDEPENDENT COMPOUNDINGSYSTEM VIEW
04 / THE LOOPFROM EXPERIENCE TO IMPROVEMENT
Turn experience into tested improvement.
Observation, proposal, comparison, adoption and restoration remain distinct stages. Candidate changes are independently evaluated before they influence future work.
FIG.03 / GOVERNED EVOLUTION PIPELINETRACE MODE
05 / WHEN IT LEARNSTHREE RHYTHMS
A micro-reflection after each exchange
A short pass over what just happened: a fact worth keeping, a preference you stated, a correction you made. Cheap enough to run every time, small enough to stay out of your way.
When the world disagrees with its expectation
Before an action it can write down what it expects to see. A mismatch is a surprise, and surprise is what triggers a deeper reflection. Calibration is measured, so confident-and-wrong is caught, not hidden.
Dreaming: trajectories become skills
In the background it re-reads recent work, merges duplicate memories, retires stale ones, and turns a repeated successful trajectory into a candidate skill. Candidates go to the loop above; nothing is promoted by dreaming alone.
06 / WHAT ACCUMULATESINSPECTABLE ASSETS
What it learns stays visible.
Memory, skills, tools, specialists, policy and evaluation results remain inspectable assets you can open, test, version, share, disable or restore. They move with your work when you change models.
Memory
Durable facts, preferences, project context, and learned constraints, each with a scope and an optional expiry.
Skills
Fifty-four skills are built in. New skills use the same format and are checked for when and how they should be used.
Capabilities
Plugins, connectors and in-process tools add new ways to work, introduced through explicit review and your permissions.
Policy
Rules include self-tests that help detect problems early and present them clearly.
Specialist agents
Persistent identities with their own brief, memory, tools, permissions, and lifecycle. Colleagues, not disposable personas.
Selection evidence
Comparable cases, regression results, your decisions, attribution, and reversal history, in a ledger you can replay.
07 / EVALUATIONINDEPENDENT TESTING BEFORE ADOPTION
Improvements should
stand up to testing.
Candidates are independently tested and compared before adoption. Evidence, not confidence, determines what carries forward.
Isolated evaluation
Candidate skills and rules are compared on representative tasks in an evaluation environment separated from your working data.
Regression checks
A test used to validate a fix reproduces the original issue, confirms the change, and remains for future work.
Evidence ledger
Comparable cases, results, decisions and restoration history remain together in a record you can review.
Isolated evaluation
A candidate skill or rule is compared against the baseline on the same tasks inside an evaluation host that cannot touch your real data. The verdict is written to the ledger before anything is applied.
Self-curriculum
It mines your own git history for tasks it once did wrong, replays them two ways, and turns the difference between the failed and the real fix into a lesson. The curriculum grows from your work, not from a benchmark.
08 / THE HORIZONRESEARCH VISION
From personal node
to a network of verified capability.
The longer-term vision is a distributed network where people exchange evaluated skills, tools and experience while deciding what to trust and adopt. This network remains in research. Private memory stays local by default, and sharing requires authorization.
09 / THE SWITCHESTWO TOGGLES · BOTH YOURS
Evolution runs only as far as you let it.
Two independent toggles in the app. Everything they produce is visible in the memory and skills panels, where any item can be disabled or rolled back with one action.
Organize memory in the background
Consolidates memory and proposes skills during idle time. A separate approval setting controls whether new skills are enabled automatically.
Approve improvements automatically
When off, every new skill waits for your review. Turn it on only when you want evaluated candidates to be adopted automatically.